This change simplifies the pre-compilation representation of the
regex to store a pre-order traversal of the syntax tree in an array
instead of an indirect enum. This enables an optimization pass that
can index into and mutate the tree more easily.

This change includes tests that verify that list-based compilation
generates the same instructions as the original tree-based
compilation, and switches to the list-based compilation. This
change also eliminates the "literal wrapper" node that preserved
AST – we aren't using the saved AST, so it resulted in unneeded 
links in the chain.

Because parsing still generates an AST which is converted to a
DSLTree before the new list, regex compilation may be slower
until the intermediate DSLTree is fully removed.

Also tweaks the RegexBenchmark formatting/settings a bit.

This optimization recognizes quantifications followed by
an atom that can't be matched at the start of the quantification,
and converts them to possessive, eliminating any backtracking
at that position. This change includes an accessor for the
first "required" atom (which will be used in other
optimizations), and converts DSL.QuantificationKind to a
struct to support the possessification changes.

When deciding on exclusion during auto-possessification, awareness of
the current matching options is important for correct analysis. For
example, /a+A/ can be auto-possessified, but the case insensitive
pattern /(?i)a+A/ cannot be.