fix(benchmark): pyright and ruff cache clearing targets wrong paths

[w-martin]

Problem

benchmarks/benchmark_checkers.py clears these directories before each timed run to simulate a cold-cache analysis:

cache_dirs = [
    target_dir / ".mypy_cache",
    target_dir / ".pyright",
    target_dir / ".ruff_cache",
    target_dir / ".ty",
]

Two of these paths are incorrect, meaning the benchmark does not actually get a cold cache for those tools.

pyright

pyright writes its analysis cache to ~/.cache/pyright/ (macOS/Linux), not to .pyright/ within the project directory. .pyright/ in the project is only pyright's config/exclusion tracking. The timed runs therefore benefit from a warm cache, systematically underestimating pyright's cold-start time.

ruff

ruff's cache lives at ~/.cache/ruff/ (the XDG platform cache dir), not .ruff_cache/ in the project. The correct approach is to pass --no-cache to the ruff invocation rather than deleting a local directory that may not exist.

Fix

1. pyright: add Path.home() / ".cache" / "pyright" to cache_dirs
2. ruff: remove .ruff_cache from cache_dirs; add --no-cache to the ruff command in the ToolInfo definition
3. ty: verify whether ty's cache is in the project dir or in the platform cache dir for the pinned version — adjust accordingly
4. After correcting, re-run benchmarks; pyright numbers in particular may change

Additional issue: npx pyright startup overhead

The pyright invocation goes through npx, adding ~100–300ms of Node.js/npm resolution overhead to every timed run. If pyright is installed globally (npm install -g pyright) the direct binary invocation should be used instead to measure only pyright's analysis time.