[PERF] Avoid repeated path matching when evaluating dependency rules

Problem

On a generated project with 5,005 Python files and 13,738 import edges,
DependOnFileCondition.check() reuses the in-process graph as intended, but
gather_depend_on_file_violations() still scans every projected edge and calls
matches_pattern() repeatedly. The matcher normalizes or splits path labels
and runs a regex for each call. A suite of rules therefore spends most of its
time evaluating the same paths again.

In the comparison benchmark (ArchUnitPython 1.5.0 at commit
09c9d199dec095aaa60be227bfd7b5c2b7ffb107, Python 3.13.4, Windows 11),
the published fresh-process median for one direct boundary rule at 5,000 leaf
modules was 3.59 s. A separate exploratory phase profile showed:

These phase timings are one-off diagnostics, not publication-grade speedup
claims. The ten rules varied source and target folders and ran in one process.
The graph cache was hit for every additional rule; at 5,000 modules the cache
lookup totaled about 0.003 s, while rule evaluation took about 17.8 s.

Proposed investigation and implementation

1. Add a repeatable 1/10/50-rule benchmark linked to [Test] Benchmark for ArchUnitPython #55, including cold
   command latency, warm suite latency, p50/p95, and peak memory.
2. Measure the number of unique source/target labels and repeated matcher
   calls. Preserve arbitrary compiled-regex and glob behavior.
3. Match each unique label against each filter only once within a rule;
   normalize or split path labels once. Consider indexing outgoing edges by
   source if the simpler change does not remove enough work.
4. Keep any cache bounded to a check or explicit suite context. Avoid a global
   cache that retains paths from unrelated projects or becomes stale after
   edits.

Acceptance criteria

• Existing behavior and violation evidence remain unchanged for glob and regex
  filters, positive/negative rules, Windows paths, and empty-test protection.
• Clean and seeded-violation benchmark corpora still produce the same results.
• Report before/after median, p95, and memory for 1,000 and 5,000 modules,
  including 1/10/50-rule suites. The rule-evaluation phase should improve
  materially without a correctness or memory regression.