feat(experimental): add deterministic GPU graph communities

[ibgreen-openai]

Goals

• Complete the explicitly requested GPU-resident community-exploration capability with deterministic, clearly documented label propagation.
• Distinguish densely connected communities inside a single weakly connected component without claiming Louvain, Leiden, or modularity optimization.

Changes

• Export LuGraphLabelPropagation and its typed props exclusively from the optional @luma.gl/experimental/lugraph entry point.
• Initialize stable vertex-ID labels and run bounded, synchronously separated majority-voting passes over weak neighborhoods.
• Require reverse adjacency for directed graphs; use symmetric forward adjacency once for undirected graphs.
• Include exactly one self vote, ignore self-loop edges, count duplicate/reciprocal original edges independently, and resolve every support tie with the lowest stable label.
• Preserve source chunks and caller-owned outputs, allocate only one graph-owned snapshot, respect physical-buffer alias isolation, and stay within the portable eight-storage-binding limit.
• Publish optional GPU-resident final-iteration convergence, preserve isolated vertices, and fail closed on the selected forward/reverse topology overflow.
• Scale dispatch using the existing bounded multidimensional helpers, including real synthetic three-dimensional execution.
• Document actual deterministic heuristic semantics, O(sum(degree²)) work per iteration, bounded nonconvergence, and explicit non-equivalence to Louvain/Leiden.
• Add comprehensive constructor/ownership Node tests and real Chromium/WebGPU CPU-oracle coverage, including two bridged dense cliques that remain distinct communities although weak components merge them.

Verification

• nvm use: passed; Node v22.22.1.
• yarn install: attempted; unpublished dev-tools dependencies return HTTP 403. Used an independent APFS clone of existing compatible dependencies without lockfile changes.
• yarn lint fix: passed after implementation; 1,576 files checked.
• yarn build: passed across all modules, including the optional luGraph ESM/CommonJS/types exports.
• Focused Node community validation: 47 passed.
• Focused actual Chromium/WebGPU community CPU oracles: 25 passed.
• Complete Node suite: 1,073 passed, 1 skipped on the final propagated head.
• yarn test: passed on the final stacked head; 1,073 Node tests and 1,734 actual Chromium/WebGPU browser tests passed (25 browser tests skipped).
• yarn website:build: passed on the final stacked head.
• (cd website && yarn build): passed on the final stacked head.
• yarn examples:typecheck: all 46 example workspaces passed.
• yarn bundle-size: all seven production bundle ceilings passed.

Risks

• Synchronous neighborhood majority voting is a bounded heuristic and may oscillate or remain unconverged; the optional real GPU scalar reports only the final fixed-point result.
• Worst-case per-iteration work is O(sum(degree²)); this is intentionally not Louvain, Leiden, or modularity optimization.
• Directed inputs need reverse CSR; required topology overflow invalidates community outputs instead of silently returning incomplete results.