Bonsai: add radial and helical arrays to bim.add_array - #9483

Extends the parametric array with a RADIAL layer type that sweeps copies
about a centre and axis, optionally climbing along that axis to form a
helix (spiral stairs, helical ramps, corkscrew ramps, curved tiers).

Placement math is consolidated into tool.Array.child_matrix, which both
tool.Model._regenerate_array_body and ArrayPreviewDecorator now call. The
two previously carried separate hand-written copies of the offset formula,
so the drag-time ghost preview could drift from what Finish built.

New BBIM_Array.Data keys, all .get-defaulted so layers written before
radial existed place byte-for-byte as before, and bpy.ops.bim.add_array()
with no arguments is unchanged:

    type, angle, rise, rise_method, axis, center, full_circle,
    rotate_children

Notes on the design:

- The sweep is never wrapped or clamped. Multi-turn is a first-class case,
  so unlike every other ANGLE property in model/prop.py the angle carries
  no min/max — only a soft range to tame mouse drag. The damage from a
  clamp is subtle: wrapping the final angle*i product is harmless (rotations
  are periodic), but normalising the stored sweep before DISTRIBUTE divides
  it changes the step (540/16 = 33.75 degrees becomes 180/16 = 11.25) while
  the LAST instance still lands correctly, so end-position assertions cannot
  detect it.

- angle and rise take their DISTRIBUTE decisions independently. A spiral
  stair is normally known as one total and one per-copy quantity -- the
  floor-to-floor height is fixed by the building while the tread angle is a
  design choice, or vice versa -- and coupling the two modes makes that
  ordinary specification unstatable. An absent rise_method inherits method.

- Whichever half of each pair the user types, the panel shows the other:
  tool.Array.resolved_rise / resolved_angle return (per_copy, total) and
  the edit body renders the derived one. The riser implied by a
  floor-to-floor total is what building code actually constrains and it
  exists nowhere else, so without this it is only discoverable by rebuilding
  the array and measuring it. A total sweep past one turn is reported with
  its turn count, since the bare angle is exactly what hides the wrap.

- The pivot is stored relative to the source instance, never as an absolute
  world point. Children track the parent through the BBIM_Array CHILD_OF
  constraint, so an absolute pivot would stay behind when the parent moved
  and tear the ring away from its own contents.

- Switching a layer to radial seeds a pivot one bounding-box width off -X.
  Left on the object's own origin, every copy would spin in place on top of
  the original and the array would read as broken -- the same failure the
  linear path avoids with its bbox-derived offset.

- Every new AddArray property is SKIP_SAVE. Blender replays an operator's
  last-used values on the next interactive invocation and the Array panel's
  "+" button passes no arguments at all, so without it a single radial add
  would make every later "+" click silently produce a radial array.

- The display gap counts (angle_gap_count / rise_gap_count) are separate
  from the DISTRIBUTE divisors, which floor at 1 to keep the division safe.
  For a readout that floor would invent a phantom step on a lone instance.

Centre picking: bim.input_cursor_center_array takes the pivot from the 3D
cursor (Shift also takes the axis from the cursor's Z, which Shift+RMB
aligns to a surface normal), and bim.center_array_on_selected puts it on
another selected object's origin. The panel branches on type, the three
linear offset arrows hide in radial mode, and a type toggle joins the
gizmo row. Angle and centre gizmos are a follow-up.

Tests: 44 numeric cases on the placement and resolved-value functions, 5
AST forward-compat guards pinning the single-placement-site and no-clamp
contracts, and 6 BDD scenarios. One test checks the derived readout against
what child_matrix actually places so the display cannot drift from the
geometry. Tolerances are 1e-5 because mathutils is single-precision.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>