feat(experimental): add batch-preserving luDF joins and lookups

[ibgreen-openai]

Goals

Expose documented unique-right-key uint32 inner joins and bounded left lookups through the optional luDF API while preserving independently streamed left/right source batches, explicit ownership, stable row identity, and GPU-resident diagnostics.

Changes

• Add immutable, precisely typed LuDataFrame.innerJoin, LuDataFrame.lookup, and corresponding filtered/derived query methods with explicit key names and bounded index/output capacities.
• Compose the independently published preserved-batch GPUBatchHashIndex prerequisite with existing GPU hash join/query primitives; build one index from independently chunked right tables without concatenation, repacking, CPU row access, or hidden submission.
• Preserve discontinuous source-row offsets and every empty left/right batch, exclude explicit nullable keys, and retain both owned source leases until compiled results are destroyed.
• Publish stable matched left/right row identifiers, exact required counts, capacity-limited published counts, explicit per-batch overflow, source-aligned lookup masks/probes, six-word index diagnostics, and a GPU-native uniqueness/reserved-key/overflow contract-violation flag.
• Suppress all matches for duplicate right keys, valid reserved keys, or incomplete indexes; reject incompatible categorical dictionaries, invalid formats/layouts, unsupported capacities, and unsafe cumulative probe bounds before GPU work.
• Reuse one caller-owned command graph across ordered parameter updates and preserve visualization-ready selection masks without CPU readback.
• Add focused planning and real-WebGPU coverage for mismatched [3,0,5]/[2,0,3] topologies, nullable rows, discontinuous offsets, bounded outputs, duplicate/reserved/overflow contracts, repeated encodes, schema-only right tables, and source destruction.

Dependencies

• Stacks on stable sorting/top-K PR feat(experimental): add stable luDF sorting and top-K #2959.
• Merges the independently reviewable preserved-batch hash-index prerequisite PR feat(experimental): build shared GPU hash indexes from preserved batches #2960 without changing its source.

Verification

• nvm use: Node v22.22.1.
• yarn install: not rerun; reused a verified dependency tree because the existing enterprise alpha registry dependency returns HTTP 403.
• Official @vis.gl/dev-tools alpha lint configuration and final yarn lint fix passed.
• Final yarn build passed.
• Final yarn test: Node 621 passed/1 skipped; real Chromium/WebGPU 1,590 passed/25 skipped.
• Focused luDF Node suite: 8 files/71 tests.
• Focused real-WebGPU dataframe joins: 5 cases; combined preserved-batch index and join integration: 9 cases.
• yarn website:build and (cd website && yarn build) passed.
• yarn examples:typecheck and all seven yarn bundle-size budgets passed.

Limitations

• Supports unique-right unsigned scalar inner joins and source-aligned bounded lookups only.
• Many-to-many, outer, multi-key, string, and signed-key joins remain intentionally unsupported.
• 0xffffffff remains reserved; valid reserved right keys explicitly invalidate the join.