fix: correctness fixes; performance regressions

[pocketken]

A performance and allocation pass on top of 4.5.0, plus new zero-allocation span APIs and a parallel polyfill. No upstream H3 API changes; cell and index results remain identical to 4.5.0 (the projection fix below shifts geometry by a few ULP to match libh3 exactly).

Fixes 🔧

• the spherical inverse projection (ToLatLng, GetCellBoundary and its vertices, directed-edge boundaries, and everything built on them) now matches upstream libh3 v4.5.0 bit-for-bit. v4.5.0 replaced the spherical-law-of-cosines inverse with a 3D-vector construction (tangent basis + linear combination + normalize); the port now follows it in the same operation order. Center and boundary coordinates shift by a few ULP from earlier 4.5.0 builds
• fixes a codegen regression on net10.0 in GetDirectNeighbour/GetDirectNeighbourWithoutRotations: over-aggressive inlining of the generated per-digit rotation tables gave the method a large zero-initialized stack frame, making grid disk, grid ring, IsNeighbour, directed-edge and polyfill operations several times slower under the net10 JIT than under net8. net10 is now at parity with net8 on these paths (net8 was unaffected)
• Fill now traces the polygon boundary in cells before flooding inward, matching upstream polygonToCells seeding. Thin/narrow features, holes and antimeridian-crossing shapes that are contained but only reachable through non-contained neighbours (and so were previously dropped from a single interior seed) are now returned. Locked to libh3's exact cell set by a differential corpus (thin sliver, concave L, box-with-hole, antimeridian quad, disjoint multipolygon)

Enhancements 🎉

• new zero-allocation span/buffer-fill overloads across the traversal, hierarchy and set APIs: GridDisk, GridDiskDistances, GridRingUnsafe, GridPathCells, GetChildrenForResolution, CompactCells, UncompactCells and GetCellBoundaryVertices gain overloads that fill a caller-owned Span<T> and return the number of cells written, with sizing helpers MaxGridDiskSize, MaxGridRingSize, GridPathCellsSize, CellToChildrenSize and UncompactCellsSize. Allocation-free on a warm ArrayPool; the streaming IEnumerable APIs are unchanged and produce identical results
• new ParallelFill for large polygon fills: shards the polygon's envelope into horizontal strips, fills each concurrently and unions the result. Produces the same (unordered) cell set as Fill. It is opt-in and only pays off on large fills: it trades the sequential fill's flat allocation for wall-clock, and below a few thousand output cells the setup cost makes it slower than Fill

Performance 🚀

• the forward and inverse spherical projections (FromLatLng/ToLatLng, cell boundaries, areas, edge lengths) drop a large amount of redundant transcendental work while staying bit-for-bit with the reference: precomputed per-face-center sin/cos and axis-azimuth tables, angle-subtraction identities in place of per-call atan2/sin/cos, a cancelled cos(latitude) in the longitude solve, a collapsed planar-radius sqrt, precomputed M_SQRT7 powers, integer round-div-by-7 on the aperture-7 up-scaling chain, and per-vertex trig cached across the Cagnoli area loop
• aperture-7 digit-pair down-steps in the resolution walk are fused into a single normalization, and ToFaceIJK fuses the aperture-7 down-step with the neighbour walk; GetDirectNeighbour fast-paths the common no-base-cell-crossing tail and skips the identity rotation reorient
• polyfill flood fill replaces NTS's per-call allocating point-in-area locator with an allocation-free inline ray-crossing locator, replaces its HashSet/Stack working sets with pooled open-addressed structures presized from the geometry's area, and hoists per-cell invariants out of the neighbour walk
• grid disk, cell boundary and edge length drop iterator and intermediate-array allocations via eager span fills and stackalloc buffers; CompactCells uses a presized bucket and a primitive ulong sort, and UncompactCells uses a pooled open-addressed dedup table and an inlined child walk

Testing

• geometry parity is gated per operation against authoritative libh3 v4.5.0: per-cell outputs within 1 ULP and the whole-sphere cell-area sum within 3 ULP, with a persisted manifest recording the worst per-op ULP observed (index and cell outputs match exactly). Added adversarial Vec3d edge-case coverage and per-op ULP diagnostics