perf(coreml): cache compiled models by default (faster runs + fixes restricted $TMPDIR) (#129)

[raphaelsty]

Fixes #129. Co-authored with the issue reporter @RBozydar.

What & why

CoreML compiles the ONNX model into a CoreML bundle at session creation. With no cache dir, ONNX Runtime compiles into the ephemeral $TMPDIR, so:

1. the model is recompiled on every colgrep invocation (each run is a fresh process), and
2. on macOS setups where $TMPDIR (under /var/folders/.../T) is rootless-restricted, the compile fails outright (the reported crash).

This points CoreML at a persistent per-user cache dir by default — ~/Library/Caches/next-plaid/coreml (honoring XDG_CACHE_HOME) — so the compiled model persists across runs and never touches $TMPDIR.

Benefit (benchmarked: per-invocation --force-gpu CoreML search, fresh process each run)

model	main (recompiles each run)	this PR (cached)	saved	speedup
LateOn-Code-edge	~1.01 s/run	~0.70 s/run	~0.31 s	~31%
answerai-small	~1.85 s/run	~1.16 s/run	~0.68 s	~37%

The first (cold) run matches main; every subsequent run loads the cached compiled model. Since searches run constantly, this compounds.

And it fixes the original crash by construction: CoreML no longer compiles under $TMPDIR, so the rootless-restriction failure can't occur.

Behavior / compatibility

• This is a deliberate default change: CoreML now writes a persistent cache under ~/Library/Caches instead of using ephemeral temp. (Opted into for the speed + robustness win.)
• Precedence: explicit NEXT_PLAID_COREML_CACHE_DIR / colgrep settings --coreml-cache-dir → default cache dir → $TMPDIR (only if the cache dir can't be created).
• Non-CoreML builds and the success path of other providers are unaffected.

Override (persistent colgrep setting)

colgrep settings --coreml-cache-dir /some/other/dir   # relocate the cache (persists)
colgrep settings --clear-coreml-cache-dir             # back to the default location

Validation

• Compiles + clippy clean with and without --features coreml; unit tests (coreml_cache_dir set/get/clear/serialize) + full suites green via ci-quick.
• Benchmarked main vs this PR on two models (table above) — consistent ~31–37% faster repeated CoreML loads.
• Smoke-tested: explicit dir + default both cache the compiled model; success path otherwise unchanged.
• Note: I could not reproduce the reporter's exact rootless-/var/folders crash locally, but caching-by-default removes the $TMPDIR dependency that causes it. @RBozydar — confirming on your environment would be the final check.

Closes #129