fix(experimental): scale visibility workflow dispatches

[ibgreen-openai]

Goals

Remove the one-dimensional WebGPU dispatch ceiling from GPU mask composition, stable compaction, and visibility workflows used by large visualization datasets.

Changes

• Reuse the existing guarded bounded three-dimensional dispatch planner and invocation indexing for GPUMask, GPUCompaction, and GPUVisibilityWorkflow.
• Propagate one device or internal synthetic dispatch limit through mask composition, source-row identity generation, hierarchical scan, and stable compaction scatter.
• Preserve public APIs, source batch/chunk topology, stable row identities, ownership, empty-selection count clearing, and the original one-dimensional behavior for smaller workloads.
• Add Node coverage for exact 65,535-workgroup boundaries, synthetic three-dimensional overflow, limit propagation, and empty inputs.
• Add a real-WebGPU regression using 1,025 rows and a synthetic dimension limit of two; verify actual [2, 2, 2] dispatches and exact selected rows/counts without allocating a giant browser buffer.

Verification

• nvm use — Node 22.22.1, matching .nvmrc.
• yarn install — blocked by enterprise-registry HTTP 403 for the existing upstream alpha development-tool dependency; reused the previously verified ignored workspace dependency tree.
• yarn lint fix — passed across the complete repository after the final source changes.
• yarn build — passed across every package after final formatting.
• Focused Node coverage — synthetic dispatch and LuxFilter checks passed; existing GPU-primitives Node coverage passed.
• Focused real-WebGPU coverage — three browser files and 35 tests passed, including the synthetic three-dimensional dispatch regression.
• CI=1 yarn test — Node: 547 passed and one skipped; real Chromium/WebGPU: 1,554 passed and 25 skipped.
• yarn website:build — passed, including validation of 461 documentation pages.
• (cd website && yarn build) — passed, including validation of 461 documentation pages.
• yarn examples:typecheck — passed across all 46 workspaces.
• yarn bundle-size — passed all seven existing bundle budgets.

Risks and compatibility

• Public APIs and GPU buffer ownership are unchanged. Dispatches below the device's one-dimensional limit retain their previous shape.
• Internal synthetic-limit helpers are source-only test hooks; oversized workloads exceeding the complete three-dimensional device capacity still fail explicitly.
• This prerequisite benefits existing visibility consumers independently of the optional luDF entry point.