feat(experimental): add live CPU and WebGPU graph benchmarks

[ibgreen-openai]

Goals

• Provide explicit-start, trustworthy CPU-versus-real-WebGPU graph benchmarks for every delivered luGraph analysis and layout path.
• Keep benchmark-only code completely outside the production experimental root and main luGraph runtime bundles.

Changes

• Add isolated optional @luma.gl/experimental/lugraph/benchmarks ESM/CommonJS/types exports and a reusable runLuGraphBenchmark(device, options) API.
• Generate deterministic sparse, dense, scale-free, disconnected, and high-degree graph datasets while preserving original nonempty/empty/nonempty source edge batches.
• Implement independent CPU oracles for directed CSR, bounded BFS with deterministic predecessors, weak components, normalized dangling-aware PageRank, exact force integration, and genuinely approximated near/far spatial forces.
• Execute the matching six actual GPU implementations, reject incorrect outputs before reporting measurements, and wait on real GPU completion fences for every timed submission.
• Report observed minimum/median/p95/max CPU times, encoding overhead, fence-synchronized GPU timings, optional genuine hardware timestamps, imported/transient memory, and approximation error without inventing missing measurements.
• Expose actual GPU weak-component convergence, real compiled iteration budgets, and the actual final PageRank L1 residual without fabricating convergence timing or adding GPU submissions.
• Time source upload, graph compilation, explicit correctness readback, and a separately fenced standalone GPU spatial-index rebuild independently; expose actual caller-owned index memory.
• Add an SSR-safe, explicit-click LiveBenchmarkPanel with all five graph families, selectable sizes, graceful missing-WebGPU behavior, and updated accurate spatial-acceleration API documentation.
• Add Node isolation/dataset/CPU-oracle regressions plus actual Chromium/WebGPU execution of all five dataset families, all six GPU algorithms, and independently fenced index builds.

Verification

• nvm use: passed; Node v22.22.1.
• yarn install: attempted but registry returns HTTP 403 for unpublished/uncached dev-tools dependencies. Used an isolated APFS copy-on-write clone of an existing compatible dependency tree; no lockfile/dependency changes.
• yarn lint fix: passed after final changes; 1,566 files checked.
• yarn build: passed, including distinct dist/lugraph/benchmarks ESM/CJS/type outputs.
• Focused Node benchmark/CPU-oracle tests: 44 passed.
• Focused actual Chromium/WebGPU tests: all five graph families passed across all six GPU algorithms.
• Complete pre-commit Node suite: 1,019 passed, 1 skipped.
• yarn examples:typecheck: passed across 46 example workspaces.
• yarn bundle-size: all seven production bundle ceilings passed.
• yarn test: passed on the final propagated head; 1,019 Node tests and 1,707 actual Chromium/WebGPU browser tests passed (25 browser tests skipped).
• yarn website:build: passed on the final propagated head.
• (cd website && yarn build): passed on the final propagated head.

Risks

• Timings are explicitly browser/device-specific; they are never presented as general performance guarantees.
• Dense and exact-force workloads intentionally remain bounded because they can be quadratic.
• Spatial output is independently validated against its actual CPU approximation; the separately reported error compares that approximation with exact forces.