Summary

Follow-up to #493, completing the two items it deferred. adaptive-triangulation 0.3.1 fixes the known degenerate-simplex divergence (curvature losses crashed on collinear input with the Rust backend), and 0.3.0 added two batched APIs targeting LearnerND's remaining Python hot loops. This PR bumps _MIN_RUST_VERSION to (0, 3, 1), adopts both APIs, and adds the deferred Rust-backend CI job.

Changes

• Batched tell_pending: distributing a pending point over the simplices that contain it previously looped tri.point_in_simplex over every simplex sharing a vertex with the containing simplex (~14 checks per point in 2D, ~70 in 3D, ~96% misses). With the Rust backend this is now one tri.simplices_containing(point, simplex=hint) call (feature-detected via hasattr, so a directly-passed custom triangulation class keeps working). The pure-Python backend keeps the original loop, refactored into _simplices_containing_point + _add_pending_point_to_simplex so both paths share the insertion code.
• Rust default_loss: when the Rust backend is active, LearnerND uses adaptive_triangulation.default_loss as the default loss_per_simplex — the same embedded-simplex-volume computation without the per-simplex Python tuple-splicing. Exposed as triangulation_backend.rust_default_loss (None on the python backend) to avoid a circular import with learnerND. Pickling adds no new constraint: a learner on the Rust backend already references adaptive_triangulation objects.
• CI: new pytest_rust_backend nox session (installs .[test,other,rust], runs the suite with ADAPTIVE_TRIANGULATION_BACKEND=rust) and a test-rust-backend job in nox.yml.
• rust extra bumped to adaptive-triangulation>=0.3.1; _MIN_RUST_VERSION = (0, 3, 1) (0.2.x installs now fall back to the python backend rather than using a version without the degenerate-simplex fix).

Performance

End-to-end LearnerND vs adaptive 1.5.0, both with the Rust backend (best of 3):

The gain grows with dimension (vertex stars get bigger). Sampled points are identical between the python and rust backends — asserted by a new test that runs both for 200 points.

Testing

• Full suite with the Rust backend forced: 294 passed, 0 failures (the 7 curvature-loss failures documented in Add optional Rust-accelerated triangulation backend for LearnerND #493 are gone with 0.3.1).
• Full suite without adaptive-triangulation installed: 292 passed (rust-specific tests skip).
• New tests: test_rust_default_loss_matches_backend (both env configurations) and test_rust_backend_samples_identical_points.
• pre-commit (ruff, ruff-format, mypy) passes on the changed files.