Improved vtkCellGrid evaluation performance and API

vtkDGEvaluator::ClassifyPoints has been rewritten for substantially better performance.

Classification now runs in parallel over cell batches via vtkSMPTools.

Before invoking the Newton-Raphson solver, each candidate probe point is tested against the convex hull of the cell’s corner points using the new vtkConvexHull static API; points outside the hull are rejected immediately, avoiding the expensive Newton step for the vast majority of probe points in practice.

The separate EvaluatePositions pass has been merged into ClassifyPoints, so parametric coordinates computed during classification are stored directly, eliminating a redundant second pass. Divergence detection now uses the new GetSignedParametricDistance method instead of a second virtual dispatch, and the linear solve uses partialPivLu instead of HouseholderQR, which is sufficient and cheaper for the guaranteed 3×3 Jacobian.

A bug where ClassifyPoints crashed when a connectivity array’s component count exceeded the number of cell corners has also been fixed.

vtkDGInterpolateCalculator::Evaluate and EvaluateDerivative now use thread-local scratch arrays for intermediate computations, enabling safe concurrent evaluation without per-call heap allocation.

All vtkDGCell subclasses now implement GetSignedParametricDistance(rst), which returns the signed distance of a parametric point from the reference-element boundary (negative inside, zero on the boundary, positive outside). IsInside(rst, tol) is now a non-virtual inline wrapper around it.

vtkStaticPointLocator::FindPointsWithinRadius (and its 2D counterpart) is now faster. Buckets lying entirely outside the query sphere are rejected early via a distance-to-AABB test, and buckets lying entirely inside it skip the per-point distance check and bulk-copy their point IDs directly. Both optimizations also apply to vtkStaticPointLocator2D.