chore: add self-benchmark suite under benchmarks/

[chad-loder]

Summary

• Adds a yarlpattern-only micro-benchmark suite under benchmarks/ covering the four hot paths: constructor cost, URLPattern.test(), URLPattern.exec(), and the yarl.URL input fast path.
• 23 benchmarks across 4 files. No production-code changes, no behavior changes.

Why a separate suite (not in tests/)

The benchmark suite lives outside tests/ so the default pytest invocation does not pick it up (testpaths = ["tests"]). It is contributor / maintainer tooling — not something end users need to run.

Layout

benchmarks/
├── conftest.py                 shared fixtures + 4 representative pattern shapes
├── bench_construction.py       7 construction benchmarks
├── bench_test.py               6 test() benchmarks
├── bench_exec.py               6 exec() benchmarks
└── bench_yarl_fast_path.py     4 string-vs-yarl.URL pairs

Pattern shapes cover the spectrum: literal-only, named-group, regex-constrained, wildcard-tail, and the MDN kitchen-sink multi-component dict.

How to run

just bench           # run, print summary
just bench-save      # run, save baseline keyed on HEAD SHA

The harness is gated behind a PEP 735 [dependency-groups] entry (bench), pinning pytest-benchmark>=5.1.0. Not exposed as a published extra.

Test plan

• just bench — all 23 benchmarks pass (locally verified)
• uv run pytest -q — 580 passed, 19 skipped (no regression — default pytest does not touch benchmarks/)
• just lint — all tools green (ruff, mypy, pyright, ty, semgrep, etc.)
• Signed commit (ED25519)

Scope notes

Deliberately self-comparative only. Cross-library URLPattern benchmarks have their own methodological pitfalls (different parser, different regex engine, different I/O) and are out of scope for this PR.

Part of the v0.2.0 roadmap.