feat(experimental): build shared GPU hash indexes from preserved batches - #2960

Goals

Build one reusable GPU hash index from streamed, independently owned record batches so hash joins can support multi-batch right-hand tables without implicit concatenation, repacking, CPU readback, or command submission.

Changes

• Add public GPUBatchHashIndex, typed properties, and batch-aware storage statistics over ordered GraphVectorView<'uint32'> key chunks.
• Support aligned per-chunk payload vectors or explicit per-batch stable source-row offsets, plus optional GPU validity masks.
• Clear the shared table and diagnostics once, then insert/finalize each source chunk in order while preserving globally earliest duplicate selection and accumulating six GPU-resident statistics.
• Exclude ordinary nulls silently; report valid reserved keys, duplicate valid keys, probe overflow, and bounded probe statistics explicitly.
• Refactor existing single-batch GPUHashIndex build passes without changing its API or existing graph node identifiers.
• Add focused Node and real-WebGPU regressions for [2,0,3] topology, mismatched chunks, discontinuous source offsets, sliced physical offsets, nullable rows, reserved keys, duplicates, capacity overflow, empty inputs, aliases, and repeated encoding.

Verification

• nvm use: Node v22.22.1.
• yarn install: not rerun; reused the already verified dependency tree because the existing enterprise alpha registry dependency returns HTTP 403.
• yarn lint fix: passed after final changes.
• yarn build: passed after final formatting.
• yarn test: Node 550 passed/1 skipped; real Chromium/WebGPU 1,557 passed/25 skipped.
• yarn test-node: 550 passed/1 skipped.
• Focused new real-WebGPU tests: 4 passed; existing hash-index/hash-join integration: 15 passed.
• yarn website:build and (cd website && yarn build): both passed and validated 461 documentation pages.
• yarn examples:typecheck: all 46 workspaces passed.
• yarn bundle-size: all seven budgets passed.

Risks

• Reserved key 0xffffffff remains unsupported and is explicitly counted; null rows do not increment invalid-key diagnostics.
• Existing single-batch GPUHashIndex behavior and public graph node IDs are preserved.
• Source rows, buffers, and command submission remain caller owned.