fix(core): return deepcopy from get_params to prevent state mutation

[pankajbaid567]

PR Title: Fix: get_params exposes mutable nested objects (#522)

Description

This PR fixes issue #522, where GAM.get_params() exposes direct references to the Term and TermList objects through the parameter dictionary. Because of this, modifications to the returned dictionary such as params["terms"][0].lam = ... would silently mutate the underlying model's state, violating the scikit-learn estimator contract.

This is fixed by updating Core.get_params() to return copy.deepcopy() of the dictionary and nested lists/objects rather than returning direct referential dictionaries.

Changes Made

• pygam/core.py: Added import copy and wrapped returned variables from Core.get_params() in copy.deepcopy().
• pygam/tests/test_get_params.py: Introduced a new test file validating that external mutations to the dictionary returned by get_params do not change the underlying structure.

Testing

• Extended the test suite with test_get_params.py (verified it appropriately captures parameter changes without model changes).
• Ran full regression testing suite (pytest pygam/tests); all passes.
• Fully compatible with sklearn's clone logic.

Closes #522

[dkstlzk]

Nice fix and good regression test!

Also thanks for picking up the issue reported in #522.

One potential concern is the use of copy.deepcopy on all parameters.
For larger models this may become expensive since get_params() is often called inside model selection utilities.
Would it make sense to deepcopy only the nested mutable objects (e.g. terms) instead of the entire parameter dictionary?

[pankajbaid567]

Thanks for the thoughtful review, @dkstlzk ! Great point

deepcopy on every parameter is indeed wasteful, especially in model selection loops like GridSearchCV where get_params() is called frequently.
I've updated the implementation (commit d47dcc7) to use a _maybe_deepcopy()helper that only deepcopies mutable types (list, dict , ndarray, and Core subclasses like Term/ TermList).
Immutable scalars (int, float, bool, str, None) are now returned as-is with zero copy overhead.

Also added two new tests to verify:
Scalar params preserve identity (no unnecessary copy)
Mutable params like callbacks are still independent copies
Full test suite passes (165/165). Let me know if you have any further suggestions!