feat(experimental): add spatially accelerated GPU graph force layout

[ibgreen-openai]

Goals

• Add an explicitly approximate, GPU-native spatial acceleration path for progressive luGraph layouts without calling it Barnes–Hut, ForceAtlas2, or exact all-pairs repulsion.
• Preserve caller-owned renderable coordinates, explicit grid memory/capacity/status, portable WebGPU bindings, and bounded three-dimensional dispatch.

Changes

• Introduce optional-subpath-only LuGraphSpatialForceLayout, composing an existing LuGraphForceLayout with caller-owned grid offsets, stable IDs, float32 cell centers, accepted counts, overflow flags, bounds, opening threshold, and exact near-cell radius.
• Rebuild the existing GPU uniform-grid index every iteration and gather per-cell centers from grouped vertex IDs without floating-point atomics.
• Compute nearby repulsion exactly, approximate sufficiently distant cells with population-weighted monopoles, and preserve every remaining interaction; theta: 0 executes exact interactions across all cells.
• Preserve directed/undirected attraction, deterministic reset, warm starts, pinned vertices, velocity clipping, render-bindable positions, and explicit GPU synchronization.
• Fail closed on out-of-domain vertices, insufficient index capacity, or required topology overflow while preserving positions and clearing velocities.
• Add 83 focused Node validation/ownership cases and 27 real Chromium/WebGPU CPU-oracle cases, including exact-versus-approximate error, grid offsets/centroids, zero vertices, 4-byte views, reset/pinning, and 1,025-vertex/1,025-cell synthetic 3D dispatch.

Verification

• nvm use: passed; Node v22.22.1.
• yarn install: attempted but blocked by HTTP 403 for unpublished/uncached dev-tools dependencies. Used an isolated APFS copy-on-write clone of an existing compatible dependency tree; no dependency or lockfile changes.
• yarn lint fix: passed; 1,561 files checked.
• yarn build: passed after final formatting.
• Focused Node tests: 83 passed.
• Focused real Chromium/WebGPU CPU-oracle tests: 27 passed.
• Complete yarn test: 975 Node tests passed, 1 skipped; 1,701 real-browser/WebGPU tests passed, 25 skipped.
• yarn website:build: passed.
• (cd website && yarn build): passed.
• yarn examples:typecheck: passed across 46 example workspaces.
• yarn bundle-size: all seven production bundle ceilings passed.

Risks / follow-up

• This is a flat uniform-grid monopole approximation, not hierarchical Barnes–Hut or ForceAtlas2.
• Complexity is O(V × C + local pairs + E) plus one explicit index rebuild per iteration, with caller-owned O(V + C) grid storage.
• The next stacked change adds explicitly started, CPU-oracle-validated live benchmarks with separate index/upload/readback timings and approximation-error reporting.