perf(picker): accelerate large workspace file searches

[fcoury-oai]

Why

The file picker rescored every workspace file after every keystroke. In large repositories, broad path prefixes keep thousands of candidates alive and make typing noticeably sluggish.

What changed

• Reuse the previous result set when a file-picker query narrows, and invalidate that cache when the query broadens or the available files change.
• Score candidate sets of at least 1,024 files in parallel with Rayon while preserving existing fuzzy scores, tie-breakers, and stable result ordering.
• Reject paths that cannot contain the requested fuzzy subsequence before invoking the more expensive matcher.
• Add an opt-in, five-sample performance benchmark that supports both a real workspace and a portable 12,000-file fixture.
• Add focused regressions for incremental narrowing, backspacing, replaced file lists, and stable ordering on large parallel-filtered sets.

Median filtering time for the codex-rs/core/src query in an optimized build:

Workspace	Files	Before	After	Improvement
Real Codex checkout	6,415	106.7 ms	22.8 ms	4.7x
Synthetic fixture	12,000	100.3 ms	19.5 ms	5.1x

How to Test

1. Run cargo test -p red --lib ui:: -- --test-threads=1; all 345 UI tests should pass, including the large-list ordering and query-broadening regressions. The manual benchmark remains ignored during normal runs.
2. Run RED_FILE_PICKER_BENCH_ROOT=/path/to/large/workspace cargo test --release -p red --lib file_picker_large_workspace_performance -- --ignored --nocapture; the benchmark should report five-sample medians for discovery, installation, keystrokes, filtering, and drawing. Omit the environment variable to exercise the portable 12,000-file fixture.
3. Open the file picker in a large repository, type a multi-character filename or path query, then backspace and change the query; results should remain responsive, correctly ranked, and include files that become eligible again.

Also verified with cargo clippy --all-targets --all-features -- -D warnings.