AG-10819: A Custom Filter Option means the same thing wherever it is used

FIX AG-10819

FIX AG-14491

FIX AG-16738

Custom Filter Options (colDef.filterParams.filterOptions) work in the Advanced Filter, taking 0, 1 or 2
values, the same arities the column filter supports. The other two tickets are bugs found on the way there.

Breaking bug fixes

Both are TypeScript-only: corrections to types that were already wrong. What stops compiling is code the
declared type should never have accepted.

1. IFilterOptionDef.predicate is required. An option without one was offered and matched no rows.
2. type accepts a Custom Filter Option's displayKey, on ISimpleFilterModel, the nine
   *AdvancedFilterModel, and IFilterPlaceholderFunctionParams.filterOptionKey. The declared unions were
   already wrong: choosing a custom option has always stored its displayKey there, and the source cast
   as ISimpleFilterModelType to compile. Only reading type into a variable of the narrow union, or an
   exhaustive switch on it, stops compiling.

filterOptions, defaultOption and filterPlaceholder narrow only what they suggest: every value they
took before is still accepted, and a wrong one is reported at runtime instead.

What a user could notice

Everything observable in one place, each labelled with what it is.

Two degenerate hand-written models differ, neither reachable from the UI: a combined model with
conditions: [] filters on nothing rather than hiding every row under OR, and one that also needs its
filterType corrected keeps its own keys instead of collapsing to {filterType}.

The feature

A column's filterOptions are offered in the expression editor and the Builder under their displayName,
localised through their own displayKey as the column filter's dropdown label is, and evaluated with the
same predicate as the column filter. Operators resolve per column, so two columns
can reuse one displayKey for different options, and an option whose displayKey is a built-in replaces it
for that column. Suggestions carry one entry per key, as the column filter's dropdown does.

Expression syntax. An option is written under its displayName followed by as many values as it takes.
Two are comma separated, brackets optional. Each value is quoted by its Cell Data Type as for the built-ins:
number and bigint unquoted, the rest quoted. A displayKey typed in place of the displayName is
replaced by it once recognised.

[Age] Even Numbers
[Age] Between (Exclusive) (30, 40)

Builder. A two-value option renders two value pills, labelled Value From / Value To from the existing
inRangeStart/inRangeEnd locale text, and held to the same out-of-order check the column filter reports on
its inputs, and reported through that filter's own locale key: strictMaxValueValidation for number and
bigint, maxDateValidation for the date types. Changing a condition's column to one of another data type
clears the values.

Model. The Advanced Filter gains a filterTo slot, its condition model having had nowhere to put a
second value.

API added

Pre-existing bugs fixed

30 of them, each covered by a test that fails on unmodified latest.

Range validation

1. A stale range error survived a new model, Cancel, and reopening. Number and BigInt read this.state
   in a refresh() override after the base had overwritten it; Date had no refresh at all.
2. It blocked the narrower option chosen after it. The to input's error held the condition back once
   the option no longer had a to input.
3. A model with fewer conditions did not drop the extra ones while another condition held an error.
4. A one-input option was held to the range rule of the value left behind: Number and BigInt compared
   from against to whatever the option.
5. A range being fixed was discarded unless it was inRange. shouldKeepInvalidInputState matched the
   key rather than the two-value shape.
6. The Date Filter validated the wrong pair once a condition had been dropped. Its picker callbacks
   captured a condition index, and removing one from the middle shifts every one after it.

Custom filter options

7. A rejected option emptied the whole list, which was replaced by the one malformed entry. Invalid
   entries are skipped and the rest kept, making #73 unreachable, so it is removed.
8. An option supplying only the removed test was still offered, and matched no rows.
9. A built-in key and a FilterOptionDef of the same displayKey were both offered: two dropdown rows
   for one key, and selecting either highlighted the first.
10. defaultOption was returned unchecked, leaving the dropdown on a value it does not list.
11. getCustomOption resolved against Object.prototype, so a displayKey of toString returned an
    inherited member. The option map is a Map.
12. numberOfInputs was not guarded, and a null entry threw on displayKey.
13. A condition short of a value fell through to the built-in evaluation instead of filtering on nothing.
14. A filterOptions list swapped at runtime never reached the dropdown. An option added later could not
    be chosen and one withdrawn stayed selectable, showing as another option's applied value.
15. filterPlaceholder never saw a custom option: placeholders were not recomputed when the option
    changed, and a custom key has no locale entry, so filterOption arrived undefined.
16. A filter option belonging to another filter type was accepted in silence (AG-16738).
17. A malformed option went unreported on a column filtered only through the Advanced Filter, which never
    builds the OptionsFactory that reports it.

Number, BigInt and Date

18. numberFormatter blanked both inputs, its output going into an <input type="number">. It is shown
    only where the input can hold it and read it back, and the inputs are re-rendered when a colDef refresh
    replaces the formatter or parser. Text no parser reads is left as typed, so a refresh does not empty an
    input mid-keystroke.
19. A model summary kept the formatter it was built with, captured in the constructor while
    updateParams replaces filterParams.
20. The floating filter read its value back with Number(), so with an allowedCharPattern a value only
    numberParser could read reached the model as NaN.
21. The BigInt Filter never rebuilt its inputs when allowedCharPattern changed, though its own floating
    filter already did.
22. A date column with includeTime showed an empty picker. setDate wrote 2020-01-01 into a
    datetime-local input, which the browser blanks.
23. The Date Filter's preset-range cache never hit: expires was stored as a duration and tested as an
    absolute timestamp, re-running the relative-range function per row. Results were correct; the cost was not.

Advanced Filter

24. An option name another option name starts with never parsed (AG-14491). Over Or Equal resolved to
    Over, silently: the editor shows the expression, getAdvancedFilterModel() returns null, and nothing
    is filtered. Reachable with no custom options by localising advancedFilterNotContains to contains not.
25. An operator display name containing ) failed to parse, ) being a terminator. Subsumed by 24.
26. Only the first of several expression rewrites landed. Rewriting a typed key in place shifted every
    position still to be parsed. Rewrites are queued and applied last-first, against the text as typed.
27. The Builder's option list accepted a half-filled pair, and dropping a condition's column left a stale
    filterTo: both only looked at filter.
28. The Builder kept the previous column's operator name where the new column offered the same key under a
    different displayName.

Other

29. A combined model with no conditions threw: validateModel mapped over an absent array.
30. A key pressed in a header the grid held no focus position for threw:
    AbstractHeaderCellCtrl.shouldStopEventPropagation destructured focusSvc.focusedHeader behind a !,
    though it is typed HeaderPosition | null.

Two optimisations, with no behaviour to test: the column autocomplete list was never memoised, its cache field
being assigned nowhere, so every keystroke rebuilt and re-sorted the entries; and a column's filterOptions were
classified twice per keystroke, once for the operators and once for the keys they narrow to. Both are cached per
column now, and invalidated on the column events they are built from.