Support scikit-learn 1.6+ by replacing removed private APIs

[mo-arvan]

Summary

Closes #71

Removes the scikit-learn <=1.5.0 upper bound by replacing the two private APIs that broke in newer versions:

self._validate_data() (removed in sklearn 1.7) → public validate_data() via feature-detection compat layer
force_all_finite parameter (removed in sklearn 1.8) → ensure_all_finite via feature-detection compat layer
Also fixes BaseEstimator import to use the public sklearn.base path instead of sklearn.mixture._base.

What does NOT change

The private Gaussian mixture internals (_estimate_gaussian_parameters, _compute_precision_cholesky, GaussianMixture._initialize, etc.) are all forward-compatible through sklearn 1.8 — no changes needed.

Changes

stepmix/_compat.py (new) — feature-detection for validate_data and ensure_all_finite
stepmix/stepmix.py — replace removed APIs, add y= alias to score() and predict() to match fit() and pass sklearn estimator compliance checks
pyproject.toml — remove sklearn upper bound
.github/workflows/pytest.yaml — add Python 3.11–3.13 and sklearn 1.5/1.6/1.8 to CI matrix

Test results

All tests pass across sklearn 1.5.0, 1.6.1, and 1.8.0 on Python 3.12, and on Python 3.13/3.14 with sklearn 1.8.0.

[mo-arvan]

Hi, thank you for maintaining stepmix.

I'm a researcher who needs stepmix on an environment running scikit-learn 1.6+, and the upper version bound is currently preventing installation. I focused on keeping the changes minimal to make the review straightforward.

I've tested locally against scikit-learn 1.5.0, 1.6.1, and 1.8.0 on Python 3.12, and against Python 3.13 and 3.14 with scikit-learn 1.8.0. All tests pass in every configuration.

I would appreciate it if you could take a look when you have a chance. Happy to make any changes if you have concerns or suggestions.