Complete the drop-in API for neighbor-aware losses; numpy fast paths in the PyO3 layer

[basnijholt]

Summary

Two independent improvements to the Python-facing layer.

1. Drop-in API completeness (found by integration testing, not review)

LearnerND with a neighbor-aware loss (triangle_loss / curvature_loss_function(), i.e. nth_neighbors >= 1) crashes with the current release: it calls five Triangulation methods we never implemented, and relies on two behaviors only that code path exercises. This PR adds:

• get_vertex, get_neighbors_from_vertices, get_simplices_attached_to_points, get_opposing_vertices, get_face_sharing_neighbors — same semantics and exception types as the reference; get_opposing_vertices and get_simplices_attached_to_points are answered natively from the facet/vertex indexes (single lookups) instead of the reference's Python set scans
• get_vertices passes None entries through as None (adaptive feeds the result of get_opposing_vertices, which contains None on hull facets, straight back into it)
• SimplicesProxy supports binary set operators in both operand orders (adaptive computes ... - tri.simplices), delegating to a snapshot set

Pinned by cross-validation tests against the reference on random 2D/3D triangulations, exception-type parity, and an end-to-end LearnerND run with curvature_loss_function() driving 100 points through the previously uncovered path.

2. numpy fast paths

• parse_point / parse_points bulk-copy f64 numpy arrays instead of iterating them as Python objects (other dtypes and nested sequences keep the generic path)
• VerticesProxy.__array__ builds a PyArray2 directly instead of round-tripping through a list of Python tuples and numpy.array

Output types are unchanged everywhere (tuples/lists/sets), so the drop-in contract is untouched.

Benchmarks (A/B vs main, best of 5)

operation	main	this PR
np.asarray(tri.vertices) ×2000, 500-vertex mesh	0.189 s	0.004 s (50×)
constructor from numpy array, 2000×2D	0.0089 s	0.0077 s
add_point with numpy rows (4D, 600 pts)	0.410 s	0.400 s
macro insertion benchmarks (2D/3D/4D, hull-heavy)	—	unchanged

examples/adaptive_learnernd.py: 3.6× end-to-end (unchanged).

Verification

• pytest: 112 passed (incl. new cross-validation + LearnerND integration tests)
• cargo test --lib --tests: 18 passed; pre-commit clean