feat(experimental): add GPU PageRank with dangling redistribution

[ibgreen-openai]

Goals

• Add browser-native, GPU-resident PageRank node importance for interactive graph visualization and sizing.
• Correctly handle incoming links, dangling vertices, damping, probability normalization, and explicit residual reporting without CPU synchronization or floating-point atomics.
• Stack narrowly on the verified luGraph topology/traversal/component foundations in feat(experimental): compute GPU weakly connected components #2944.

Changes

• Add LuGraphPageRank and LuGraphPageRankProps exclusively to @luma.gl/experimental/lugraph.
• Implement directed incoming-CSR pull accumulation, with symmetric forward adjacency reused for undirected graphs.
• Redistribute dangling-vertex probability mass every iteration, apply configurable damping/teleportation, and explicitly normalize the resulting probability vector.
• Publish an optional caller-owned GPU-resident final-iteration L1 residual.
• Introduce reusable bounded three-dimensional float32 workgroup-tree reductions with uniform barriers; avoid the existing reduction primitive's one-dimensional dispatch ceiling.
• Reuse a single graph-owned float32 candidate workspace and one compact reduction hierarchy across dangling-mass, normalization, and final-residual reductions.
• Fail closed on required forward/reverse overflow; empty graphs clear optional residual without division by zero.
• Respect portable WebGPU storage-binding limits, reject physically aliased caller outputs, and preserve source ownership/chunk identities.
• Define the metric as unweighted even when topology retains optional edge weights.
• Add 54 focused Node contracts and 23 real-WebGPU CPU-oracle scenarios covering dangling vertices, damping zero/one, reverse pull, normalization, final residual, overflow, repeated encoding, byte offsets, and actual hierarchical [2,2,2] dispatch.

Verification

• nvm use: passed (Node v22.22.1).
• yarn install: attempted normally and with escalation; configured registry blocks 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 (1,548 files).
• Final yarn build: passed across every workspace after the final code/formatting changes.
• Focused Node PageRank tests: 54 passed.
• Focused real Chromium/WebGPU PageRank tests: 23 passed.
• Complete CI=1 yarn test: 806 Node tests and 1,649 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; LuGraphPageRank remains absent from the experimental root.
• Normal commit hooks: lint and all 806 Node tests passed.

Risks and follow-up

• Scores use native float32 and deterministic workgroup-tree reductions; expected tolerances account for floating-point rounding and unsorted incoming CSR accumulation order.
• Directed PageRank requires explicitly constructed reverse adjacency; existing edge weights are intentionally not interpreted as transition probabilities.
• The implementation never claims native floating-point atomics, subgroup requirements, float64 precision, or automatic CPU fallback.
• Stacked on feat(experimental): compute GPU weakly connected components #2944; target codex/lugraph-connected-components until prerequisites land.