Estimator conventions

Estimator conventions

Followed from scikit-learn, per TORML_CODING_GUIDELINES.md:

• Hyperparameters go in __init__ only and are stored unchanged on self.
• fit returns self.
• Learned attributes end with _ (coef_, classes_, n_features_in_).
• Validate inputs through torml.utils (check_array, check_X_y).
• get_params / set_params round-trip; clone rebuilds an unfitted copy.
• Mixins provide score (ClassifierMixin → accuracy, RegressorMixin → R²), fit_transform (TransformerMixin), fit_predict (ClusterMixin).
• Use vectorized PyTorch ops; no NumPy/SciPy/sklearn imports in library code.
• torch.testing.assert_close for tensor comparisons in tests.