Add autodifferentiable DelaunayNN mesh

[Jammy2211]

Summary

• add a public DelaunayNN mesh backed by Sibson/Watson natural-neighbor interpolation
• keep Qhull connectivity behind a stopped JAX callback while performing floating-point geometry in pure JAX
• support mapper construction, split regularization, boundary queries, overflow signalling, and package exports
• add focused numerical, autodiff, flip-continuity, historical C-parity, and reconstruction tests

Why

Barycentric Delaunay interpolation is local and fast, but its value and gradient can jump when the triangulation flips as lens mass-model parameters move. Natural-neighbor weights retain locality while remaining numerically stable across those flips.

The fixed table caps default to 32 cavity triangles and 32 neighbors. A 101-mass-model workspace audit observed maxima of 25 cavity triangles and 27 neighbors; caps 16 and 24 overflowed.

Validation

• focused DelaunayNN/Sibson tests: 8 passed
• broader focused inversion coverage: 43 passed
• full PyAutoArray suite: 948 passed, 51 skipped
• three unrelated existing PyNUFFT failures remain because that dependency calls removed scipy.linalg.pinv2
• Ruff and git diff --check pass