fix: handle None config and robust hparams in MambAttention (closes #449)

[botbikamordehai2-sketch]

What

This PR fixes a bug where loading a saved MambAttention model could fail or lose configuration state. The issue reports that load() sets model_config, preprocessing_config, trainer_config, and random_state to None, causing the model to silently retrain with library defaults. In MambAttention.__init__, if config is None (as set by restore_base_state), accessing config.pooling_method or config.shuffle_embeddings in the fallback would raise an AttributeError, breaking model reconstruction.

Fix

• Guarded against config being None by substituting with a default MambAttentionConfig().
• Extended the exception handling in hparams lookups to catch KeyError as well, making it more robust to missing hyperparameters.
• This ensures that loaded models retain the correct configuration and can be safely used for prediction, scoring, or further fitting.

Closes #449