perf(core): opt-in fast coincident-point merge (vendored OpenMP clean kernel)

[akaszynski]

Summary

Adds an opt-in fast path to vtkStaticCleanUnstructuredGrid for the exact-merge default regime, by vendoring pyvista-algorithms' clean kernel (MIT) and wiring a thin adapter into RequestData.

Stacked on #82 (fast surface). Base is feat/fast-surface-port; review/merge #81 then #82 first. Net-new here is the FiltersCore clean path + staticclean_fast op.

What it does

• Filters/Core/pvaClean.h (+ pvaHash96.h / pvaLinearProbeTable.h / pvaParallelRadixSort.h) — vendored MIT kernel (namespace pvu::clean), self-contained, excluded from the unity build.
• Filters/Core/fvtkFastClean.{h,cxx} — adapter: validates the exact-merge default regime (effective tolerance 0, no point-data averaging, no merge-by-data-array, concrete float/double grid, <2³¹ points, no polyhedra) and calls run_clean(remove_degenerate_cells=false) so cells stay 1:1 in input order.
• Hook in vtkStaticCleanUnstructuredGrid::RequestData, gated on fvtk::FastModeEnabled(). Default off → byte-exact.

Faithful-to-stock details

• Cells are kept 1:1 in input order → cell data passes through unchanged (matches stock PassData).
• Merged-point data is copied from the kernel's canonical source (source_map); tolerance-0 merge means coincident points share exact coordinates, so output coordinates are bit-identical to stock.
• Cell-type representation is preserved: a homogeneous grid keeps its implicit vtkConstantArray types (so the output's GetCellTypesArray() stays null, exactly like stock), a heterogeneous grid keeps its explicit per-cell array.

Robust cell-type acquisition

GetCellTypesArray() returns null for homogeneous grids (implicit constant types). The adapter handles this: zero-copy AOS pointer when present, else a per-cell GetCellType(i) copy. (Same fix applied to the surface adapter in #82.)

Validation

• Built cp312-abi3 manylinux_2_28 wheel on the executor; vtkFiltersCore.abi3.so links libgomp (OpenMP real, not stub).
• Confirmed the kernel genuinely engages (under EnableFast() the output point order is reordered vs the standard path, same point set) — not a silent fallback.
• Bit-exact gate: 292 passed — the 4 staticclean_fast cases match stock under the points-relaxed gate; standard cases stay byte-exact.

🤖 Generated with Claude Code

[akaszynski]

Merged to main via fast-forward (41fcfe9 clean kernel + 64fbfc1 MSVC gate fix), rebased onto the merged surface tip to drop duplicate lower-stack commits. The Windows build is green (the vendored OpenMP clean kernel compiles to the bit-exact stub on MSVC, falling back to stock vtkStaticClean). Live on main.