FIX: sklearn clone() returns empty terms list (#340)

[VanshKharb]

Fix sklearn.clone() returning empty terms list on fitted GAMs (#340)

Description

This PR resolves an issue where calling sklearn.base.clone(gam) on a fitted GAM estimator results in an identical estimator but with a missing terms list (gam.terms == ""), breaking Scikit-Learn workflows (like grid search and cross-validation) that rely on cloning properly passing initialization arguments.

Root cause:
GAM.fit() mutates string configurations like terms='auto' into TermList objects in-place. Because sklearn.clone() queries get_params(deep=False), it received the mutated state rather than original initialization parameters, and then additionally tried to deep clone the TermList dynamically because it inherits Core.get_params(), resulting in parameter destruction.

Changes:

• Store Initial Values: GAM.fit() now stores _init_distribution, _init_link, _init_callbacks, and _init_terms before applying data-dependent validation.
• Selective get_params() Override: GAM.get_params(deep=False) now explicitly restores the original instantiation parameters so clone() receives correct input, cleanly segregating scikit-learn requirements from pyGAM's internal keep_best deep=True routines.
• Bypass duck-typing: Custom __sklearn_clone__ implemented on Term and TermList to explicitly force copy.deepcopy and prevent sklearn.clone from attempting to reconstruct them as empty nested estimators.
• Internal Optimization: Removed redundant, harmful set_params() calls in _bootstrap_samples_of_smoothing and gridsearch that broke identically.
• Testing: Comprehensive regression test test_sklearn_clone_preserves_terms guarantees functionality safely across LinearGAM/LogisticGAM and customized parameters.