feat(experimental): add progressive GPU force-directed graph layout

[ibgreen-openai]

Goals

• Add progressive, browser-native GPU force-directed graph layout using position buffers that rendering can consume directly without copying.
• Support deterministic initialization, pinned nodes, configurable physics, and warm-started iterations per frame.
• Stack narrowly on the verified luGraph PageRank/adjacency foundations in feat(experimental): add GPU PageRank with dangling redistribution #2946 without claiming ForceAtlas2 or unsupported floating-point atomics.

Changes

• Add LuGraphForceLayout and LuGraphForceLayoutProps exclusively to @luma.gl/experimental/lugraph.
• Require caller-owned packed GPUVector<'float32x2'> positions with both STORAGE and VERTEX usage, plus independent packed float32x2 velocity storage.
• Compute exact tiled O(V² + E) all-vertex repulsion, symmetric incoming/outgoing edge attraction, origin gravity, damping, integration timestep, and bounded velocity.
• Separate globally synchronized force and integration passes; preserve stable read-only positions during force accumulation without graph-owned scratch or floating-point atomics.
• Require reverse CSR for directed graphs so every edge contributes symmetric physical attraction; reuse forward CSR for undirected graphs and explicitly ignore optional edge weights.
• Support caller-owned uint32 pin masks, deterministic seeded one-shot GPU reset controls, and persistent warm starts across repeated encodings.
• Preserve positions and clear velocities on required adjacency overflow; preserve pinned positions and clear pinned velocities.
• Support aligned four-byte-offset vector storage, bounded three-dimensional dispatch, uniform shared-memory tile barriers, and portable eight-or-fewer storage bindings.
• Add 77 focused Node contracts and 23 real-WebGPU CPU-oracle scenarios covering exact forces, deterministic reset, pinning, overflow, rendering-compatible ownership, velocity clipping, repeated progression, and actual [2,2,2] tiled dispatch over 1,025 vertices.

Verification

• nvm use: passed (Node v22.22.1).
• yarn install: attempted normally and with escalation; configured registry rejects unavailable dependencies with HTTP 403. Verification uses an isolated APFS-cloned compatible dependency tree and ignored Yarn state; no package/lockfile changes.
• yarn lint fix: passed (1,552 files).
• Final yarn build: passed across every workspace after final WGSL, test, and formatting changes.
• Focused Node force-layout tests: 77 passed.
• Focused real Chromium/WebGPU force-layout tests: 23 passed.
• Complete CI=1 yarn test: 883 Node tests and 1,672 real Chromium/WebGPU tests passed; existing skips only.
• yarn website:build: passed.
• (cd website && yarn build): passed.
• yarn examples:typecheck: all 46 example workspaces passed.
• yarn bundle-size: all seven bundle-size budgets passed.
• Built ESM and CommonJS package-subpath smoke checks: passed; LuGraphForceLayout remains absent from the experimental root.
• Normal commit hooks: lint and all 883 Node tests passed.

Risks and follow-up

• Repulsion is exact and intentionally quadratic; scalable approximate spatial acceleration is a separate follow-up, not misrepresented as exact ForceAtlas2.
• Directed layout explicitly requires reverse adjacency; edge weights are retained by topology but intentionally ignored by this initial unweighted spring model.
• Every workgroup invocation participates uniformly in tiled shared-memory barriers; no device-wide shader barrier, floating-point atomic, CPU readback, or implicit command submission is used.
• Stacked on feat(experimental): add GPU PageRank with dangling redistribution #2946; target codex/lugraph-pagerank until prerequisites land.