Document model inversion, and pin the constants the book quotes

[kgdunn]

Follow-up to #470, which added PLS.invert() and OPLS. Companion to the
pid-book PR kgdunn/pid-book#259, which writes the same material up as a chapter.

Documentation

PLS.invert() and OPLS landed with docstrings but no narrative page, so the
reasoning behind them lived only in the book. docs/user_guide/model_inversion.rst
covers:

• inverting a PLS model, and why the answer is a set of designs rather than one;
• walking the null space with null_space_coordinates, and what rising
  Hotelling's T-squared along that walk means;
• why inversion can need more components than prediction does, since it has to
  describe the X-space well enough to map a score back into it;
• the O-PLS decomposition and why, for a single response, its orthogonal space
  is the same linear space as the PLS null space;
• how well the null-space direction is actually determined, which for the
  cheddar-cheese example is not well at all;
• why a multivariate specification region is not the box of per-input ranges
  that summarises it.

It is linked from the hidden toctree in multivariate.rst, beside pca, pls,
tpls and multiblock.

Tests

These close a real gap rather than adding coverage for its own sake.

Most numbers in the book's worked example are printed by code blocks a reader
can run, so a change in this library would be visible there. A handful are
quoted in prose without appearing in any output, and nothing would have caught
a silent change to them:

Quantity	Value
Hotelling's T-squared limits, 95% and 99%	7.36, 12.14
SPE limit, 99%	1.60
Score norm at the direct-inversion solution	0.281
Score norm at the -1 and +1 steps	1.039
Step at which T-squared is least	-0.103
PLS first weight against its own loading	5.49 degrees
O-PLS predictive weight against its loading	0.31 degrees
PLS and O-PLS regression coefficients	agree to 1.4e-13
O-PLS split of the sum of squares in X	68.7 / 18.3 / 13.0

The tolerances match the precision the text quotes, so a failure means a
sentence in the book needs rewriting, not that a digit moved far downstream.
Two of the tests check identities rather than only values: that
||tau_DI + s g||^2 == ||tau_DI||^2 + s^2 for the steps taken, and that
w'p == 1, which is what makes p - w a plain subtraction.

The last test pins the specification-region result: of the eight corners of
the box of three ranges, exactly two predict an acceptable taste, both of those
are beyond the T-squared limit, and therefore none is an acceptable lot.

Verification

ruff check and mypy src/process_improve clean. The 26 tests in this file and
test_multivariate_opls.py pass.

The wider suite has pre-existing failures in this sandbox from missing optional
extras (seaborn, ridgeplot, and the DOE solver); they reproduce identically
on main with this branch stashed, and CI installs --all-extras.

Version is a patch bump, 1.62.0 to 1.62.1: documentation and tests only, no
behaviour change. CITATION.cff and CHANGELOG.md updated in step.

---

Generated by Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!