perf(locker): load the 3D hero preview faster

[Slush97]

Supersedes #420. Same goal, now with the follow-up fixes from the #406 viewer work and a few fixes #420 was missing, on current main.

What

Opening a hero's 3D preview was slow on both ends: the export on a cache miss (full-size textures, sequential metadata decodes, physics and attachments decoded after the model) and the load every time (full-size <img> textures decoded on the main thread at the first render, plus a second full decode of every albedo just to average it).

Main process (cache miss):

• Pose and rigged exports pass --max-texture 1024, so each texture embeds at its largest mip that fits (drifter skin: 39 MB -> 14 MB). POSE_PIPELINE_VERSION 16 / RIGGED_PIPELINE_VERSION 15 so cached full-size GLBs regenerate. The constant is PREVIEW_MAX_TEXTURE.
• readVpkEntryBytes keeps a small LRU of entry indexes instead of walking pak01's whole tree (~25 ms) on every lookup.
• Attachment and clip metadata blocks decode 8 at a time instead of one after another.
• Rigged physics and attachments decode while the model exports, not after.
• Skins whose rigged export drops every clip get a versioned marker, so later opens go straight to the static pose instead of exporting and loading both.

Renderer:

• The hero viewer decodes textures with ImageBitmap, off the main thread. The soul-container grid and import modals keep <img> (ImageBitmaps pin decoded pixels and the grid shows many models), through a per-parser GLTFLoader plugin instead of hiding createImageBitmap globally.
• GLBs are fetched with one native arrayBuffer() and parsed with parseAsync, skipping FileLoader's per-chunk progress re-streaming.
• The albedo average reads a 256 px off-thread decode of the source image instead of drawing the full albedo into a canvas on the main thread (75 to 90 percent of model load time before).
• Cloth bones write parent-first without per-write subtree updates (was O(L^2) per chain per step); bit-identical on the Dynamo fixture.
• Disposal closes ImageBitmaps, frees every texture a material holds (sheen maps were missed), and runs in a [scene] effect. Disposing in the loader cleanup left one frame drawing closed bitmaps when a toggle re-ran the load.

Measured

Per change, on the original branches (not re-run end to end on this combined branch yet):

• vpkmerge export, vanilla drifter with its 6-clip menu: 816 ms (v0.20.0, full size) -> ~135 ms (feat(model): parallelize GLB texture export and add --max-texture vpkmerge#48 at 1024; heroes with 4096 normal maps pay up to ~200 ms more for the top-mip normal probe). model clips --json: 127 -> 50 ms.
• Dev app over CDP, drifter + pak09 skin: 0.72 s cold (export + load), ~0.6 s cached, longest main-thread block ~120 ms.
• Electron 35, same GLB, load to first frame:

GLB	<img>	ImageBitmap	renderer memory (<img> / bitmap)
full size	1396 ms	760 ms	285 / 721 MB
2048 cap	1048 ms	515 ms	270 / 607 MB
1024 cap	449 ms	~300 ms	219 / 303 MB

<img> and ImageBitmap frames are pixel-identical at every cap. At the closest zoom, 1024 is slightly soft on fine print (drifter's hat card, coat damask); 2048 matches full size for ~300 MB more renderer memory.

vpkmerge

Bumps the pin to v0.21.0 (Slush97/vpkmerge#48, released via #54): fetch script version + sha256s, the Nix flake input and cargoHash (a full nix build .#vpkmerge reports 0.21.0 and accepts --max-texture). The released Linux binary runs both export shapes this PR issues (--pose --require-pose --max-texture 1024 and --clip ... --max-texture 1024).

Known, not fixed here

From review, low impact:

• With a trippy paint or the unified NPR material active, the original albedo (and maps the NPR clone nulls) live outside the scene's materials when it is disposed, so their ImageBitmaps are left to GC instead of closed.
• R3F unmounts asynchronously, so a rAF can still render the disposed scene into the detached canvas. three catches the texture upload errors; at worst it's console noise.
• The clipless marker, like the cached GLBs, is versioned by RIGGED_PIPELINE_VERSION only. A vpkmerge or game update that changes clip export needs a version bump to clear it.

Checks

tsc -b, eslint ., vitest run (183 files, 2718 tests) and ui:check pass. interchange.test.ts fails locally while Deadlock is running (it hits the real "Game is running" check); unrelated. loadGltfPreview.test.ts covers the plugin swap, the ImageBitmap opt-in, the failed-fetch path, materialTextures and disposeTexture.