perf: byte-level path handling on Unix to bypass Components iterator

[stormslowly]

Why

std::path::Components runs a full state machine on every iteration — Windows-prefix detection, Component enum construction, double-ended-iter bookkeeping. On Unix the resolver doesn't need any of that: OsStr is raw bytes and /, ., .. are always single-byte ASCII regardless of UTF-8 content in segments.

What

• src/path.rs — unix_normalize and unix_normalize_with replace Components-based PathUtil::normalize / normalize_with on Unix. Non-Unix paths keep the existing implementation.
• src/cache.rs — byte-level parent_path replaces Path::parent in Cache::value (the recursive parent walk that drove most of parse_next_component_back). realpath_uncached uses a new join_last_segment instead of Path::strip_prefix + normalize_with, since parent.path is already a strict byte prefix of self.path when produced by parent_path.

Scope change

An earlier revision of this PR also rewrote require_without_parse to dispatch on the specifier's first byte via a hand-rolled classify_specifier_head table. That hunk was extracted to #246 and closed — the maintenance cost of keeping a parallel impl of Path::components in sync with std outweighed the small isolated win.

The previous PR body cited −7.28% Ir on resolver[single-thread]; that number included the now-dropped opt3. A fresh callgrind diff for opt1+opt2 alone is pending and will be posted before this PR is taken out of draft.

Test plan

• cargo test --lib -- --skip 'tests::pnp::' — 138 / 138 pass (1 environment-only symlink fixture failure also present on main)
• cargo clippy --all-features -- -D warnings — clean
• Fresh callgrind diff on resolver[single-thread] for opt1+opt2 alone
• CI bench job to confirm the same delta in the codspeed simulator
• Windows CI for the #[cfg(not(unix))] branches

[codspeed-hq]

Merging this PR will improve performance by 5.13%

⚡ 5 improved benchmarks
✅ 5 untouched benchmarks

Performance Changes

	Mode	Benchmark	BASE	HEAD	Efficiency
⚡	Memory	resolver[pnp resolve]	8.4 KB	8.2 KB	+3.1%
⚡	Simulation	resolver[resolve from symlinks]	146.4 ms	139.1 ms	+5.21%
⚡	Simulation	resolver[single-thread]	50.3 ms	46.1 ms	+9.24%
⚡	Simulation	resolver[[multi-threaded]resolve]	57.9 ms	55.7 ms	+3.98%
⚡	Simulation	resolver[[multi-threaded]resolve from symlinks]	75.6 ms	72.6 ms	+4.2%

Tip

Curious why this is faster? Comment @codspeedbot explain why this is faster on this PR, or directly use the CodSpeed MCP with your agent.

---

_{Comparing perf/ascii-fast-path (8cefe78) with main (080188f)}