feat(gltf): support independently animated scenegraph instances

[ibgreen-openai]

Goals

• Let several scenegraph consumers independently animate the same immutable, postprocessed glTF asset without introducing a parallel skeleton, animation mixer, material system, or glTF loader.
• Make the resulting scene and animator directly usable by deck.gl-style getScene / getAnimator integrations.
• Keep independent node transforms, clip selection, phase, playback speed, crossfades, skin palettes, morph buffers, material-pointer state, and resource ownership correct on both WebGL and WebGPU.

Changes

• Add public createGLTFAnimationInstances, GLTFAnimationInstances, GLTFAnimationInstance, and typed creation/selection options to @luma.gl/gltf.
• Reuse the existing createScenegraphsFromGLTF, GLTFAnimator, engine AnimationMixer, GLTFSkinController, and canonical morph-target path for every isolated character.
• Support independently positioned roots; named source nodes; clip selection and crossfades; normalized clip phases and absolute seconds; pause/play, reverse/speed, and once/repeat/ping-pong looping; safe removal and idempotent disposal.
• Keep authored source accessors/images immutable and shared while giving every character its own mutable scenegraph, materials, mixer actions, joint matrices, and packed morph vertex buffers.
• Automatically delegate to the scenegraph-owned resource lifecycle when available, keeping the companion shared-resource-cache foundation independently mergeable.
• Add a substantial dedicated API guide, glTF documentation tab/sidebar, cross-navigation, fixture-backed regressions, and live browser rendering tests.

Verification

• Cross-package strict TypeScript project-reference build: node_modules/.bin/tspc -b modules/*/tsconfig.json --pretty false.
• 172 Node tests across 34 suites covering glTF, ANARI, engine animation, and experimental rendering.
• Dedicated SimpleSkin.gltf, AnimatedMorphCube.glb, and AnimatedColorsCube.glb regressions prove independent roots, source immutability, skin palettes, GPU buffer bytes, material pointers, clip transitions, timing, pause/reverse, and safe disposal.
• Real headless Chromium hardware Metal WebGL + WebGPU draws two simultaneous independently skinned characters and two independently morphed characters; the same coverage also passes SwiftShader software WebGL + WebGPU.
• Additional untracked local verification using CC0 RobotExpressive.glb: 14 named clips, three simultaneous Dance/Running/Walking characters, two independent 43-joint palettes per character, facial morph targets, an actual 50/50 Dance→Wave crossfade, and safe surviving-instance destruction. The external robot asset is not vendored by this PR.
• Full production Docusaurus build succeeds and indexes 474 documentation pages; generated route, tab/sidebar, and internal links verified.
• Scoped Biome formatting/lint and git diff --check pass.
• The legacy repository pre-commit script cannot run in the existing installation because injected YARN_NO_PROXY is rejected by its installed Yarn configuration; equivalent strict build, 172 tests, both native/software GPU runs, and the complete documentation build were run explicitly before committing. The normal organizational pre-push hook remained enabled.

Architecture / Ownership

• All new runtime ownership stays in @luma.gl/gltf; existing rendering, interpolation, shaders, and retained-object packages remain unchanged.
• Shared source asset data does not imply one instanced GPU draw: independently deformed poses currently remain separate render groups on both backends. Shared GPU texture caching belongs to the independently mergeable asset-resource lifecycle PR, not this manager.
• Targets master directly.