GLTFMetalKit: GPU compute morph kernel for high-vertex-count assets

## Context

GLTFMetalKit currently applies morph targets via a CPU pre-pass (`GLTFPrimitiveMorphData.morphedVertices(weights:)` in `Sources/GLTFMetalKit/Renderer/GLTFRenderable.swift`), rebuilding a fresh interleaved vertex buffer each frame and uploading via `device.makeBuffer(bytes:length:)`.

This is fine for the canonical Khronos test asset (`AnimatedMorphCube`, 24 verts, 2 targets). It will NOT scale to blendshape-driven faces (typical VRoid avatar: ~30K verts × 30 expression blendshapes), where the CPU loop and the per-frame upload become the per-frame bottleneck.

The pattern already exists in `Sources/VRMMetalKit/Shaders/MorphAccumulate.metal` (SoA layout, `morph_accumulate_positions` + `morph_accumulate_normals` + `morph_accumulate_combined` kernels). VRMMetalKit's expression system uses these. The kernels are spec-correct glTF morph math — they're VRM-tagged by location, not by content.

Surfaced in PR #241. Deferred under the rule "extract on demonstrated need, not speculation" — no GLTFMetalKit consumer has hit the limit yet.

## What's needed

- Move `MorphAccumulate.metal` (or fork) into `Sources/GLTFCore/Shaders/`.
- Add a `GLTFMorphCompute` Swift driver that mirrors VRMMetalKit's dispatch path: SoA delta buffers, active-set filtering on non-zero weights, dispatch position + normal kernels per frame.
- `GLTFAsset.drawCalls(animationIndex:time:)` picks the GPU path when `morphData.targetCount > N` (threshold tunable, maybe 4 or by total vertex count).
- Keep the CPU pre-pass for small assets — sub-millisecond CPU work is faster than a compute dispatch's overhead.

## Acceptance

- `AnimatedMorphCube` continues to render identical to the current CPU path.
- A synthetic stress test (e.g. 30K verts × 16 targets at varying weights) shows the GPU path produces correct output and stays within budget for 60 fps real-time playback.
- VRMMetalKit's existing `MorphAccumulate.metal` usage is unaffected (no breakage in the 1441 existing tests).

## Estimate

Modest. The kernels are written. The work is in the Swift driver + threshold heuristic + a stress test. ~1 day.

Start with GLTFPrimitiveMorphData.morphedVertices in Sources/GLTFMetalKit/Renderer/GLTFRenderable.swift and compare its path with Sources/VRMMetalKit/Shaders/MorphAccumulate.metal. Trace GLTFAsset.drawCalls(animationIndex:time:) and the existing VRMMetalKit dispatch usage before implementing the driver and path selection. Done means AnimatedMorphCube remains identical, the 30K-vertex stress test produces correct output within the 60 fps budget, and existing tests remain unaffected.

Written by the indexing model from the issue text.