Add SupernoteAtelier parser for Atelier .spd files

[philips-clanker]

Summary

• Adds SupernoteAtelier, a parser for .spd files produced by the Supernote Atelier app.
• Unlike .note files (a custom binary layout, parsed by SupernoteX), .spd files are plain SQLite databases (tiled canvas images + a small config table). Reads them via sql.js and exposes:
  • config — every config row as raw bytes
  • fmtVer, thumbnailBuffer, viewport (vp.x/vp.y/vp.scale), canvasSize (surface.width/surface.height), layers — decoded from config
  • surfaces — the per-tile PNG blobs from every surface_* table in the file, keyed by table name (discovered dynamically)
  • toImage(surfaceName) — stitches one surface's tiles into a single image, sized/positioned against the tile-grid bounds of every surface in the file (not just the requested one), so different layers line up and can be composited
  • toCompositeImage() — flattens every surface into one final image directly, layered bottom-to-top by layers order (best-effort). The simple one-call API for "just give me a finished picture."
• The .spd schema (especially the ls/layers config encoding) isn't officially documented. The initial implementation was reverse-engineered from https://github.com/Ziv-Ink/Atelier-parser, a community tool that writes .spd files — then checked against a real device-generated .spd file, which surfaced real bugs the community tool didn't expose:
  • Surfaces aren't limited to surface_1/surface_2. Real files can have arbitrary surface_{layerId} tables (e.g. a "Reference Layer" used surface_9999), so surface tables are discovered from sqlite_master instead of assumed.
  • Different layers can have sparse, differently-offset tile coverage, so per-surface tight-crop stitching produced images with different origins that couldn't be composited correctly. toImage now sizes/positions against the shared tile grid across all surfaces.
  • Building toCompositeImage surfaced two more, both invisible when looking at a single surface's image in isolation and only apparent once actually layering surfaces together: image-js defaults a new image's buffer to opaque black, not transparent, so unpasted grid cells were being treated as real content and wiped out whatever was underneath; and the reused compositeImages treats a pixel as transparent only when its packed RGBA is exactly zero, which a transparent-but-non-black pixel (e.g. transparent white) fails, so tiles are now normalized to zero RGB wherever alpha is zero before compositing.
  • This also confirmed the tile-id addressing scheme (tid = col * 4096 + row + offset) generalizes correctly to a real, non-default canvas size.
• Test fixtures:
  • tests/input/sample.spd — synthetic, shaped to mirror the real file (sparse per-layer tiles, a surface_9999 reference layer, an empty surface_3, surface.width/surface.height config).
  • tests/input/real-device.spd — a real Atelier export from a Supernote device (a built-in "Audubon" template background with a couple of sketch strokes on two layers), used to verify the parser against actual hardware output.

Closes #31.

Test plan

• npm run build (tsc) succeeds
• npm run lint passes
• npm test — full suite passes (33/33), including tests/atelier.test.ts against both fixtures
• Visually inspected stitched/composited PNGs from both fixtures to confirm correct tile alignment and layer stacking
• Pixel-level assertions that compositing leaves untouched areas identical to the background and does change pixels under foreground layer tiles