cca-zoo v3.0.0 — Full rewrite with cleaner API, strict typing, and new docs

Release notes:

What's new in v3.0.0

This is a ground-up rewrite focused on simplicity, correctness, and
long-term maintainability.

API changes (breaking)

• Unified sklearn-style interface across all model families —
  fit(), transform(), fit_transform(), score(), get_params(),
  set_params() work consistently everywhere
• model.weights replaces the old per-model attribute names
• JointData(n_views=...).sample() replaces the old data-generation API
• Deep models no longer ship built-in encoder architectures — pass your
  own nn.Module encoders; this removes the hidden coupling to specific
  network shapes
• Probabilistic CCA results are now in model.posterior_samples_

New / improved models

• Gradient linear models: PLS_EY, CCA_EY, MCCA_EY — Riemannian
  gradient descent on the Stiefel manifold with polar retraction
• Full multiview support throughout: MCCA, GCCA, TCCA, KGCCA, KTCCA,
  DCCA variants, DTCCA

Quality

• Fully typed codebase — passes mypy --strict with zero errors
• Passes ruff check and ruff format with zero errors
• Test suite with correctness/optimality checks (not just "code runs"):
  • CCA finds >0.95 correlation on correlated data; perfect on identical views
  • MCCA(2 views) == CCA; rCCA(c=0) == CCA; KCCA(linear, c→0) ≈ CCA
  • Canonical variates are orthogonal across dimensions
  • Gradient methods converge within atol=0.05 of exact eigendecomposition

Documentation

• New MkDocs + Material theme site at
  https://jameschapman19.github.io/cca_zoo/
• API reference auto-generated from docstrings
• Five clean example scripts covering all model families

Removed

• Built-in deep encoder architectures (architectures.py)
• Old Sphinx documentation