feat(resources): load prepared image packs on the 3DS SD worker

[doodlewind]

Apps with a complete local atlas currently stream every missed raster tile from a paired desktop and convert its pixels on the 3DS render path. Add optional io.resource-pack support so installed, prepared images load from SD through a native worker and the existing resource collection lifecycle.

• createPackedImageCollection uses one cache for an indexed local image and an optional desktop fallback. Views still own demand; cancellation, stale-response release and texture eviction remain in the framework. Disconnected read loaders refuse admission so local work can continue.
• The desktop writer prepares PICA RGB565 layout and independent zlib records. The reader bounds entries, indices, lengths, staging, file handles and frame admission; validates CRC; and fences old realms. No filesystem calls or pixel arrays enter guest JS.
• GPU-owned core texture descriptors eliminate the retained CPU pixel copy. The renderer adopts/retires storage after its GPU fence. A 256×256 tile uses 128 KiB GPU RGB565 instead of 256 KiB RGBA8; a shared one-image/mesh materialization budget remains.
• Document the pack format, installation contract, budgets and diagnostics. PSP and other unsupported hosts retain the explicit desktop fallback.

Validation: 134 core tests; 113 resource/offload/platform tests; actual C worker under ASan/UBSan (stalled reads, corrupted and incompressible entries, exhausted slots, stale tickets, realm reset); TypeScript check; full Pocket Map production and diagnostic 3DS builds. Prepared color-grid bytes matched devkitPro tex3ds 2.3.0 after applying the renderer's vertical origin.

Physical 3DS validation: Pocket Map completed four SD/Mac cold-view pairs and two eight-second diagonal pans, with the requested storage source in every leg and zero native pack failures. Mean resource-ready time was 292.75 ms from SD versus 1,177.50 ms from Mac (4.02×). During the same 320 px/s route, callbacks with at least one target-level tile not ready fell from 96.36% to 13.10%. SD pan completed 63 reads and 61 uploads; Mac pan performed no SD reads or uploads. Both binaries and the complete 354 MiB atlas passed device readback and are installed.

This is one fixed-route run with a 40-entry guest cache, the same prediction policy, annotations disabled and Mac caching enabled. Callback frequency was 54.35/s on SD versus 51.40/s on Mac; it is not a GPU presentation measurement or a sustained-60-FPS result. Target-not-ready counts do not measure blank pixel area. The connected comparison does not verify offline cold startup. Measurement method, receipts and limits.

This is stacked on Draft #377, which remains unmerged; only the SD resource work is in this diff.