perf(splines): drop the n_train square projector from tprs transform

[ChrisW09]

Addresses the transform-side half of #13. Deliberately does not close it — see below.

Problem

ThinPlateSplineTransformer.transform rebuilt the null-space projector on every call:

P_new = np.eye(Z.shape[0]) - Z @ ZTZ_inv @ Z.T   # n_train x n_train
K_new_proj = K_new @ P_new

Z.shape[0] is the training size, so the cost of transforming a few rows scaled
quadratically with the fit — 250 MB and 42 ms to transform ten rows against an 8k-row fit —
and the pinv was recomputed each time.

Fix

The dense projector is never needed. Since P = I - Z (Z'Z)^-1 Z':

K_new @ P  ==  K_new - (K_new @ Z) (Z'Z)^-1 Z'

(Z'Z)^-1 is now cached at fit time as _ztz_inv_.

Result

Numerically identical — max absolute difference 1.4e-14 against the explicit-projector
formulation over a 300-row fit and 37 query points including out-of-range ones.

n_train	transform 10 rows, before	after	peak alloc before	after
2000	6.3 ms	0.41 ms	—	0.64 MB
8000	42.4 ms	0.72 ms	256.7 MB (at n=4000)	2.56 MB

Scope — why this does not close #13

#13 also covers the fit side, where a dense n x n cdist and a full eigh make 8k rows
take ~40 s and 50k rows get OOM-killed. That cost is inherent to the exact TPS
construction, so fixing it means either documenting a supported ceiling, raising above a
threshold, or implementing a knot-subsampled low-rank variant (Wood's thin-plate regression
splines). Each is a product decision rather than a bug fix, and adding a size guard would
introduce a new user-visible failure mode.

I have left #13 open for that decision and marked this Refs, not Closes.

Tests

Three added to tests/test_thinplate_transformer.py:

• test_tprs_transform_matches_explicit_projector — pins numerical equivalence against a
  literal np.eye-based projector, so the algebraic rewrite can't silently drift
• test_tprs_transform_does_not_allocate_a_train_sized_matrix — tracemalloc peak must stay
  under a quarter of a dense n_train² float64 array; this fails on the old code
• test_tprs_caches_the_pseudo_inverse

Full suite: 483 passed, 9 xfailed. ruff check clean on changed files; pyright unchanged at
its pre-existing 71 errors.

🤖 Generated with Claude Code