Add PLS model inversion (null space) and O-PLS support

[kgdunn]

What and why

Adds latent-variable model inversion and an O-PLS estimator to
process-improve, motivated by García-Carrión et al., "On the Equivalence
Between Null Space and Orthogonal Space in Latent Variable Regression Modeling"
(Journal of Chemometrics, 2025, e70057). The paper proves that, for a single
response, the null space from PLS model inversion and the orthogonal space
from O-PLS are the same linear space. This PR brings both capabilities into the
package; it also underpins a companion page in the pid-book.

PLS.invert() - model inversion + null space

Given a desired response on the original Y scale, PLS.invert():

• returns the minimum-norm (direct-inversion) input vector the model predicts
  will achieve it;
• returns an orthonormal basis for the (A - r)-dimensional null space (the
  family of input vectors that all yield the same prediction);
• accepts null_space_coordinates to move along that space and meet secondary
  criteria (cost, safety, operability) without changing the prediction;
• reports the solution's Hotelling's T2 to flag extrapolation.

Verified against the paper's Case Study 1 (Table 3).

OPLS - Orthogonal PLS estimator (single response)

Trygg-Wold O-PLS NIPALS: one Y-predictive component plus
n_orthogonal_components Y-orthogonal components. Methods: fit, transform
(predictive scores), correct (orthogonal-signal-corrected X), predict, and
invert. OPLS.invert() reaches a desired response through a single division
and returns the orthogonal-space basis.

Tests confirm the paper's results on real and synthetic data:

• O-PLS(1; A-1) reproduces A-component PLS predictions and beta to ~1e-14;
• orthogonal scores are Y-orthogonal (To' y = 0);
• the O-PLS orthogonal space and the PLS null space span the same subspace
  (principal cosine = 1.0).

Worked example (cheddar cheese)

The cheddar-cheese dataset (already used elsewhere in the pid-book PLS section)
is vendored as an offline fixture. A test trains on cheeses 5-30 and designs a
chemistry toward each held-out cheese's taste; every design round-trips to its
target, and the more ambitious targets land at higher T2 (a teaching point for
the book).

Housekeeping

• Version 1.58.0 -> 1.59.0 (new public API); CITATION.cff and CHANGELOG.md
  updated in step.
• New tests: test_multivariate_inversion.py, test_multivariate_opls.py, plus
  OPLS added to the sklearn-compatibility checks. ruff and mypy clean; no
  regressions in the multivariate suite.
• CI fix (not caused by this diff): CI's uv sync had auto-upgraded ruff
  past the lockfile to a version that enabled new rules (CPY001, RUF100/ISC
  changes), reddening lint on every PR. Upper-bounded ruff<0.16 in both dev
  groups (mirroring the existing mypy pin) and added an explicit CPY001
  ignore for the repo's house copyright style.

Follow-up (separate repos)

• A pid-book page in the PLS section tying the cheese example to the
  null-space / orthogonal-space equivalence, with a figure in kgdunn/figures.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!