perf(files): optimize file search delete batches

[JustYannicc]

Summary

• Collapse nested watcher delete paths before tombstoning the file-search index.
• Replace the entries x delete-prefix scan with direct-id tombstones plus descendant prefix/ancestor-set matching.
• Add a focused synthetic benchmark and regression coverage for file deletes, directory deletes, nested delete collapse, and non-matching prefix-like paths.

Benchmark

Command: node scripts/benchmark-file-search-delete-batch.mjs

Synthetic scenario: 68,162 index entries, 4,041 nested delete paths, 8,041 tombstoned entries.

• Before: 13,295.52 ms
• After: 21.42 ms final run (observed after runs: 16.41 ms, 15.85 ms, 25.38 ms, 21.42 ms)
• Improvement: about 621x faster on the final run

Validation

• node --test scripts/test-file-search-delete-batch.mjs
• node --test 'scripts/test-*.mjs' (passes; live Electron test skipped because Electron is not launchable here)
• Codex LSP diagnostics on src/main/file-search-index.ts: no errors
• pnpm run build:main
• pnpm run check:i18n exited 0; it still reports pre-existing Italian locale gaps for settings.ai.whisper.vocabulary and settings.extensions.appSearchScope, and this PR does not touch user-facing strings

Risks

• Delete matching remains string/path based and preserves the previous case-sensitive behavior.
• Tombstoned entries remain in pathToEntryId as before, so re-added paths can clear the deleted flag through indexEntry.

[JustYannicc]

Superseded by the consolidated upstream PR #535: #535