Core data-structure overhaul: interned simplices, facet adjacency, incremental hull

[basnijholt]

Summary

Reworks the core triangulation storage for performance and maintainability. The Python API is unchanged and drop-in compatible with adaptive.learner.triangulation; all 64 cross-validation tests pass unmodified.

Commit 1 mechanically splits the PyO3 surface (argument parsing, error conversion, proxies, PyTriangulation) out of triangulation.rs into a new src/py.rs, leaving the core pure Rust.

Commit 2 replaces the FxHashSet<Vec<usize>>-everywhere layout — which cloned and re-hashed whole simplices in every hot path — with interned storage:

• a slab assigns each simplex a u32 SimplexId; the sorted vertex list is stored once and referenced by id everywhere else
• vertex_to_simplices becomes vertex → FxHashSet<SimplexId> (integer sets instead of sets of cloned Vecs)
• a new facet index maps every (dim−1)-face to its incident simplex ids, maintained by a single link_simplex/unlink_simplex pair, together with the derived boundary-facet set (the convex hull) and an overfull-facet count (corruption detector)

This addresses three known hot spots:

• locate_point walk steps and containing() facet queries are now one hash lookup (previously a set intersection over cloned Vecs)
• the bowyer_watson cascade finds neighbours via dim+1 facet lookups instead of unioning and scanning every simplex around all cavity vertices
• extend_hull reads the maintained boundary set instead of recounting every face of every simplex per outside-hull insertion — this was the O(n²) bottleneck for hull-heavy workloads — and hull() is now O(hull)

reference_invariant() additionally cross-checks the facet index, boundary set, and overfull count against a recount, so the stress suites verify the incremental bookkeeping.

Benchmarks

A/B against unmodified main, same machine, best of 3, incremental insertion (add_point loop):

workload	main	this PR	speedup
interior 2D, 5000 pts	0.296 s	0.143 s	2.1×
interior 3D, 2000 pts	0.540 s	0.163 s	3.3×
interior 4D, 500 pts	1.959 s	0.241 s	8.1×
hull-heavy 2D, 3000 pts	3.271 s	1.058 s	3.1×
hull-heavy 3D, 1500 pts	3.395 s	0.698 s	4.9×

examples/adaptive_learnernd.py end-to-end: 3.5× vs pure-Python LearnerND (unchanged from main).

Verification

• pytest: 64 passed (cross-validation vs Python reference on randomized 2/3/4D inputs)
• cargo test --lib --tests: 18 passed (incl. new tests for facet-index consistency, incremental hull, and containing facet queries)
• Stress sweeps (50 seeds × {mixed-scale, off-origin small-scale, near-duplicates, random} × {2D, 3D}): identical pass/fail profile to main (failures only on the inputs documented in src/tolerances.rs where the Python reference fails more often; robustness work is a separate follow-up PR)
• pre-commit clean