fix: ensure LogisticGAM complies with scikit-learn classifier contract

[KaranSinghDev]

Description

This PR addresses the label-handling bug reported in #578 . It enables LogisticGAM to behave as a standard scikit-learn classifier by auto-encoding labels and exposing required metadata.

Changes

1. Auto-Encoding: The fit method now uses np.unique to safely encode any binary labels (strings, booleans, etc.) into 0/1 integers before passing them to the core _pirls engine.
2. Expose classes_: The fitted model now has a classes_ attribute containing the original label names, as required by scikit-learn.
3. Predict with Original Labels: The predict method now decodes the binary output back into the original label space (e.g., predicting ["spam", "ham"]).
4. Validation: Added explicit logic to raise a ValueError if the data is multi-class (>2 unique labels).

Testing

I have included a new test suite and confirmed that all 162 existing tests pass locally in a clean environment.

I am open to any reviews or feedback you may have regarding this implementation. and would like to hear interested people's thoughts on this.