Fix Scikit-Learn API compliance and GridSearchCV support (#340, #482)

[VanshKharb]

Summary

This PR resolves several compatibility issues with Scikit-Learn estimators, specifically addressing #340 (sklearn.clone attribute mutation) and #482 (GridSearchCV failures). Following these changes, pyGAM models now pass the full sklearn.utils.estimator_checks.check_estimator suite.

Detailed Changes

• Fix estimator cloning (Returning no terms when cloning the model #340): Updated GAM.get_params() to return the original un-mutated init arguments. Previously, standard string arguments were replaced with objects during fit(), which caused sklearn.clone() to fail.
• Fix GridSearchCV compatibility (LogisticGAM not usable with GridSearchCV #482): LogisticGAM.fit() now defines the classes_ attribute, and predict() correctly maps outputs to these class labels instead of returning boolean masks.
• Input validation improvements: Added targeted ValueError and TypeError exceptions in check_array(), check_X(), and check_y() to correctly handle 1D arrays, dense requirement for sparse matrices, and zero-feature arrays based on Scikit-Learn guidelines.
• Immutable defaults: Converted mutable default list arguments in init (e.g., callbacks) to tuples.
• Estimator state properties: Added n_iter_ tracking to GAM._pirls to satisfy iterations checks.
• Binary validation: Added explicit multiclass rejection to LogisticGAM.

Testing

• Added tests/test_sklearn_gridsearch.py to verify end-to-end GridSearchCV compatibility with LogisticGAM.
• Confirmed all existing unit tests in pygam/tests pass.