feat(experimental): add dense luDF group aggregation

[ibgreen-openai]

Goals

Add GPU-native dense categorical grouping to the optional luDF dataframe while preserving immutable query planning, explicit null semantics, source batches, graph ownership, and direct visualization-ready output.

Changes

• Add precisely typed frame.groupBy(key).aggregate(definitions) and composable filtered/derived query variants for unsigned dense category keys.
• Reuse GPUGroupAggregation, GPUMask, and the existing dataframe command graph for count, sum, min, max, and mean across all original source batches.
• Infer complete dense domains from explicit categorical dictionaries or accept an explicit positive group count; preserve ordered output aliases, category labels, and source metadata.
• Exclude null category keys; count all remaining rows while floating-point statistics additionally exclude null, NaN, and infinite values.
• Publish explicit per-statistic GPU validity masks, distinguish empty groups from zero-valued sums, cache shared input masks, and keep every result resident in owned GPU buffers.
• Preserve per-batch row masks/indices/counts, reusable encoder-ordered filter parameters, derived-column validity, empty source tables, caller-owned submission, and source destruction leases.
• Add focused Node/API contracts and real-WebGPU coverage for mixed empty batches, all five metrics, null keys/values, NaN/infinity, repeated encodes, derived inputs, empty dictionary domains, ownership, and exact CPU-equivalent results.

Verification

• nvm use — Node 22.22.1, matching .nvmrc.
• yarn install — enterprise registry returns HTTP 403 for an existing upstream alpha dev-tool dependency; reused the previously verified ignored dependency tree.
• yarn lint fix — passed across the complete repository after final source changes.
• yarn build — passed across every workspace package after final formatting.
• Focused Node: all existing and new luDF suites — 46 tests passed.
• Focused real WebGPU: dense grouping, ordered parameters, derived values, nonfinite filtering, and schema-only sources — five tests passed.
• CI=1 yarn test — Node: 589 passed and one skipped; real Chromium/WebGPU: 1,570 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.

Scope and risks

• Initial groups require dense uint32 keys and float32 numeric statistics; signed dictionary indices, sparse keys, and CPU-side domain inference are intentionally unsupported.
• Counts are unsigned 32-bit and reject source row counts beyond that representable range; floating-point aggregation follows the existing WebGPU primitive's float32 precision and explicitly rejects nonfinite contributions.
• Group result rows are globally merged into one dense output batch; original source selection masks and stable source-row identities remain batch-aligned.
• Global reductions, numeric histograms, stable sorting, joins, and integrated examples remain separate follow-on changes.