feat(gltf): reference-grade asset lifecycle, shared resources, and complete API docs

[ibgreen-openai]

Goals

Make @luma.gl/gltf a reference-quality, explicitly owned asset runtime instead of repeatedly uploading source images and leaving scenegraph GPU resources alive after assets are replaced.

Keep loading in loaders.gl, glTF interpretation in @luma.gl/gltf, generic scenegraph primitives in @luma.gl/engine, shading in @luma.gl/shadertools, and ANARI as a thin retained facade.

Changes

Deterministic asset lifetime and shared GPU resources

• Add public GLTFAssetResourceCache, scoped to its creating device and keyed by source image identity, canonical sampler state, source color space, and color/data interpretation.
• Reuse compatible feature-specialized parsed materials across source materials, primitive layouts, variant mappings, and independently created scenegraphs.
• Add public GLTFScenegraphs.destroy() / dispose() to release unique models, materials, hidden scenegraph nodes, detached source-mesh templates, generated textures, and all four GPU-instance attribute buffers exactly once.
• Add explicit ParseGLTFOptions.resourceCache sharing with deferred destruction and textureCount, materialCount, and referenceCount diagnostics.
• Preserve caller ownership of the luma device, source images, explicit shared caches, and supplied image-based-lighting textures.
• Export the previously inaccessible public ParseGLTFOptions type.

Accurate reference documentation

• Document every returned scenegraph property, source-native material variants, camera/light/visibility pointers, automatic skin palettes, strict extension validation, shared-cache ownership, and descriptor-based interchange.
• Correct obsolete glTF and ANARI claims about source GPU instancing, recursive visibility, automatic retained skeletal animation, morph/camera pointers, and animated glTF export.
• Clarify that standalone glTF scenegraphs retain their established transmission fallback while experimental/ANARI scene rendering owns captured-scene refraction.
• Replace unqualified “lossless” positioning with source-faithful descriptor-based interchange in package metadata and the module overview.

Verification

• node_modules/.bin/tspc -b modules/anari/tsconfig.json --force --pretty false — passed, including strict glTF, engine, effects, experimental, and shader project references.
• Focused glTF, ANARI, exporter, showcase-identity, and capability suites — 26 suites / 134 node tests passed.
• Dedicated lifecycle coverage proves image deduplication, distinct sampler/color interpretations, variant reuse, deferred shared-cache destruction, borrowed IBL preservation, hidden-node disposal, and exact texture/buffer count restoration.
• Real Chromium native-Metal WebGL + WebGPU regression matrix — 6 suites / 28 GPU tests passed.
• The same dedicated lifecycle matrix on CI-style Chromium SwiftShader — passed on WebGL and WebGPU.
• Full Docusaurus production build — passed; llms normalization and validation passed for 474 raw documentation pages.
• All changed MDX pages compile; equivalent scoped Biome formatting/lint and git diff --check pass.
• Three-way merge audits against the simultaneous independent-animation and next-PBR branches report zero conflicts.
• Standard yarn lint fix / yarn build wrappers cannot run against this workstation’s pre-existing outdated @vis.gl/dev-tools installation (ocular-lint / ocular-clean missing); equivalent direct Biome and complete referenced TypeScript build both passed. Existing PushPatrol pre-push protection was preserved.

Architecture / Ownership

This PR targets master directly and does not depend on another unmerged branch. The cache and lifecycle remain format-owned; no loader, BRDF, animation mixer, high-level renderer, or independent ANARI material implementation is introduced.