fix(experimental): scale GPU sort dispatches

[ibgreen-openai]

Goals

Remove one-dimensional WebGPU dispatch limits from stable GPU bitonic/radix sorting while preserving the existing public sort and batch-sort APIs.

Changes

• Reuse the existing guarded three-dimensional workgroup planner and invocation linearization across bitonic initialization, compare/exchange stages, final gather, radix classification, and radix scatter.
• Propagate one device or internal synthetic dispatch limit into every nested radix GPUScan stage.
• Preserve stable equal-key ordering, direction, padded bitonic networks, source offsets, out-of-place ownership, empty/single-row fast paths, and GPUBatchSort chunk preservation.
• Reject impossible padded workloads before mutating graphs or allocating scratch resources.
• Add focused synthetic-device Node coverage and actual-WebGPU tests using only 1,025 rows and a dimension limit of two to validate [2, 2, 2] workgroups for both sorting algorithms.

Verification

• nvm use — Node 22.22.1.
• yarn install — enterprise registry blocks the existing alpha development-tool dependency with HTTP 403; reused previously verified ignored dependencies.
• yarn lint fix — passed after final source changes.
• yarn build — passed across every package after final formatting.
• Focused Node sorting and existing primitive coverage — passed.
• Focused actual WebGPU sorting — eight tests passed, including bounded 3D bitonic/radix dispatch and stable duplicate keys.
• CI=1 yarn test — Node: 549 passed and one skipped; actual 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 example workspaces.
• yarn bundle-size — passed all seven existing bundle budgets.

Risks and compatibility

• Existing public APIs, sort direction/stability guarantees, physical buffer ownership, and one-dimensional workloads remain unchanged.
• Synthetic dispatch-limit entry points remain internal; bitonic padded-capacity overflow and workloads beyond three-dimensional limits fail explicitly before partial graph mutation.
• This independent prerequisite benefits every existing GPU sort consumer and unblocks large-dataframe sorting.