Add a persistent, size-bounded rasterization cache

[philips-clanker]

Summary

• Adds RasterCache, a content-addressable cache of rasterized page PNGs keyed by (hash of the .note file's own bytes, page number), with two tiers:
  • Disk tier: persistent, capped at 100MB, LRU eviction. Storage is a minimal CacheStorage structural interface satisfied by Obsidian's app.vault.adapter, so the module has no dependency on the obsidian package and is unit-testable with a plain mock.
  • Memory tier: a smaller (20MB default), purely in-memory LRU layer in front of the disk tier, checked first. Never persisted — starts empty every plugin reload — but skips disk I/O + base64 decode entirely for pages already served earlier in the same session.
• Wires the cache into ImageConverter.convertToImages: pages already cached are served without touching the worker pool; only missing pages are rendered and written back.
• Three perf bugs found and fixed while validating this against real, larger notes (the raster cache alone wasn't moving the needle on perceived open/close speed):
  1. ImageConverter used to own and tear down its own WorkerPool on every rasterization call, and WorkerPool's constructor eagerly spawns hardwareConcurrency new Web Workers — so even a 100%-cache-hit open (zero pages rendered) still paid full multi-worker spin-up/teardown for nothing. Now a lazily-created, module-level singleton reused for the plugin's lifetime, torn down once in onunload().
  2. SupernoteView.onLoadFile awaited createImageBitmap() one page at a time inside its per-page DOM-building loop — a page-count-scaling cost paid on every open regardless of how the PNG was obtained. Now decoded up front via Promise.all.
  3. The big one: SupernoteView builds an in-memory PDF purely so pdf.js can extract a text layer for search/selection (pdfPage.render() is never called against it — the embedded image is never actually shown). It was still unconditionally embedding every page's real image into that PDF. pdf-lib's PNG embedding always fully decodes and recompresses the source PNG at .save() time regardless of it already being compressed — real, size-proportional cost. Profiled against an actual user-provided 12-page note with several user-uploaded-background pages: ~2.9s / ~4.5MB → ~40ms / ~43KB, same extracted text either way. Fixed by bumping the supernote-typescript submodule to add addTextOnlyPdfPage() (philips/supernote-typescript#34) — a PDF page with the recognized-text layer but no image — and using it for this internal-only pipeline. The user-facing "export/attach as PDF" commands are untouched and still embed real images, since those PDFs are genuinely opened by the user.
• Reuses the existing non-cryptographic hashBytes fingerprint from deviceSync.ts rather than adding a new hashing dependency.
• One shared RasterCache instance is created in SupernotePlugin.onload(), stored under the plugin's own data folder (<plugin dir>/rasterCache/) — never as visible vault content.
• Adds a "Rasterization cache" row to the settings tab (disk + memory tier stats, "Clear cache" button) and debug-level console logging (cache init, per-open hit/miss/timing, page-render phase timing, background PDF pipeline timing) — this instrumentation is what let the above three bugs get root-caused with real numbers instead of guesswork.
• Every storage operation is fail-open: a read/write/parse error is logged and treated as a cache miss or no-op, so a broken/unavailable cache degrades to "always re-rasterize" instead of breaking rendering.

Design notes:

• Cache granularity: keyed by the whole note's content hash + page number, not a per-page hash — editing any page invalidates cache entries for every page in that note. Chosen for simplicity; stale entries just age out via eviction.
• Eviction policy: LRU on both cache tiers via Map insertion-order reordering on hit; the disk tier's reordering is kept in memory only, persisted on the next put()/clear() rather than on every read.

Closes #129.

Test plan

• npx tsc -noEmit -skipLibCheck — clean
• npm run lint — clean (pre-existing unrelated warnings only)
• npx vitest run — 109 passed, including 14 rasterCache.test.ts cases (roundtrip, per-key isolation, LRU eviction on both tiers, memory tier avoiding repeat disk reads, independent per-tier eviction budgets, index persistence, fail-open behavior, clear())
• supernote-typescript's own test suite (26 tests, vitest) passes with the new addTextOnlyPdfPage, including a new test asserting identical extracted text vs. addPdfPage at <1/10th the byte size
• ./scripts/build — full production build succeeds
• Manually verified real cache population/persistence, and the three perf fixes, against actual user-provided notes during testing