Correct load_class bugs in tabsurvey models

[joelromanky]

Summary:

This PR updates the load_class function in torch_models.py to ensure that model scalers are correctly re-initialized during loading. It now automatically loads ScalerData from a scaler.json file if one is present, ensuring the model's internal dimensions align with the training state.

Changes:

• Modified load_class in tabularbench/models/torch_models.py to check for and load scaler.json if a scaler is not explicitly provided in kwargs.
• Ensured TabScaler is properly fitted with the loaded ScalerData before model initialization, preventing size mismatch errors in the state_dict during load_state_dict().
• Added a validation check to raise an error if scaler.json is missing, ensuring configuration integrity.
• Updated model-related logic in the tabularbench/models/tabsurvey folder to ensure consistency with the new loading flow.

Impact:

• Resolves RuntimeError: Error(s) in loading state_dict (size mismatches) that occurred when loading pre-trained models.
• Ensures that models correctly reconstruct their input/feature dimensions from saved training data, rather than defaulting to empty/incorrect sizes.