Price a mesh cage by the control points that were dragged

[leonardoaraujosantos]

Why

mesh::Lattice::displacement summed every control point at every vertex: nx * ny * nz multiply-adds whatever had been dragged. At the 32³ ceiling that is 32,768 terms per vertex. ClaySpace previews a mesh cage drag by laying the cage over the mesh on every pointer move, and one frame with a single corner in hand measured:

cage	per frame, before (ClaySpace, 62k vertices)
3³	~10 ms
8³	~35 ms
32³	~1.7 s (4.7–5.2 s on the 80k-triangle fixture in CyberdyneCorp/ClaySpaceDesktop#176)

The offset field is linear in the offsets, so a control point at rest adds exactly nothing. That whole cost was work with no result.

What changes

• Lattice keeps the set of control points with a non-zero offset, maintained by set_offset (a point set back to zero leaves it). displacement sums over that set only; is_identity() is that set being empty (O(1)); dragged_count() reports its size.
• Each axis's Bernstein basis is built in O(n) from powers of t and 1 − t with precomputed binomials, in double, rather than de Casteljau's O(n²) recurrence in float — negligible beside the old n³ sum, not beside the new one.
• BM_MeshLatticeDrag times one preview frame (apply with a record, revert) at 3³, 8³ and 32³ with one point dragged over ~100k triangles.
• OpenSpec change price-a-mesh-cage-by-its-dragged-points adds the meshing requirement; docs/07 states the cost.

No C ABI, Python or file-format change. The SDF lattice deformer (clattice_point) is a separate evaluator and is untouched.

Evidence

• New unit tests in test_mesh_lattice.cpp: the sparse sum matches the full trivariate Bernstein sum (reference written the slow way, de Casteljau in double) at 2, 3, 8 and 32 divisions, inside, on and outside the box; the corner of a 32³ cage still interpolates exactly; the dragged count follows writes including a point returned to rest. Existing lattice tests unchanged and passing.
• ctest --preset cpu-only: all unit suites pass locally (the one local failure, clay_device_bench_selftest, is the host's Python 3.9 rejecting float | None in tools/check_device_bench.py; unrelated).
• BM_MeshLatticeDrag (CPU time, loaded machine): 3³ 2.9 ms, 8³ 3.7 ms, 32³ 8.8 ms.
• End to end in ClaySpace with this engine: a single-point drag frame at 32³ went from ~1.7 s to ~11 ms on a 62k-vertex mesh (3³ and 8³ ~9–10 ms, the remainder being normals and the host's own work).

Needed by CyberdyneCorp/ClaySpaceDesktop#176; ClaySpace consumes it once it ships in a release.