Compile member-selectors once instead of per resolution

[ndreno]

Stacked on #1 (base is that branch; the diff here is only the perf change). Merge #1 first, then this rebases onto main.

The lever

A regex member-selector (REQUEST_HEADERS:/^X-/, !REQUEST_COOKIES:/__utm/) was kept as a string and recompiled with the regex crate on every value it was tested against, on every request. CRS carries 162 rules with a !REQUEST_COOKIES:/__utm/ or /_pk_ref/ exclusion, and the exclusion closure recompiles the selector once per resolved cookie, per rule, per request.

Measured in isolation, that recompilation alone:

cookies= 1   selector evals=  162   recompile-each= 1.359 ms   cached= 0.004 ms   saved 99.7%
cookies= 5   selector evals=  810   recompile-each= 3.251 ms   cached= 0.014 ms   saved 99.6%
cookies=20   selector evals= 3240   recompile-each=12.519 ms   cached= 0.061 ms   saved 99.5%

Against the ~2 ms the gateway measures for a whole inspection, this was the dominant cost, hiding in the exclusion path. It also corrects an earlier conclusion in this project's notes ("borrowing bought only 7%; the ~2 ms is genuine rule evaluation") — the recompile, not evaluation, was the bulk of it on requests with cookies.

The change

Compilation now turns each parsed Target into a CompiledTarget whose regex selector is a compiled regex::Regex, built once. Resolution (Variables::resolve, push_map, the exclusion retain, count_of) matches against the compiled form directly and rebuilds nothing. grep confirms no regex::Regex::new remains in the resolve path (was 2 call sites).

The parsed Target/Selector in the AST are untouched, so the sealed rule-set format is unchanged — only the in-memory compiled program gains the new CompiledTarget/CompiledSelector types, which is where the "compile the AST into an evaluable program" design says a compiled automaton belongs.

This also subsumes the selector validation from #1: compiling the selector is what validates it, so check_targets became compile_targets and an uncompilable selector still fails the build with the same InvalidSelector error. Both #1 tests still pass through the new path.

Verification

• 169 unit tests pass; clippy and fmt clean.
• Real CRS 4.9.0 still compiles to 590 rules, same 4 detectSQLi/detectXSS refusals, zero unexpected errors.
• Exclusion semantics preserved through the compiled path: a request carrying __utmz=attack is allowed by a rule that session=attack trips.
• Both fail-closed behaviors from Refuse invalid selectors and backward skipAfter at compile time #1 survive: invalid selector refused, backward skipAfter refused.
• The FTW regression job is the end-to-end detection guard.

Not included

The evaluate_step per-value to_vec() copy (a second, smaller perf lever) and the loose whole-collection exclusion prefix are still open follow-ups from the review.