Fixes to GraniteGuardian metric,, safety evals cleanups

[bnayahu]

• Granite Guardian metric to properly handle data_classification_policy
• Interim fix to make 'prediction' available to the metric.
• Further cleanup to safety evals.

[martinscooper]

@bnayahu @elronbandel note the the following added lines makes to impossible to set an input field named 'prediction' as it will be overwritten:

# TODO replace with logic inside verify_granite_guardian_config and process_input_fields
task_data["prediction"] = prediction

In addition, we should discuss how the fix should be implemented. Using the predictions only makes sense, conceptually, in the case where the risk is evaluating a generated response (the response). So it would be all the assistant risks and some of the RAG test cases.

What promoted this change was to be able to mix LLMAsJudge metrics with Guardian metrics. LLMAsJudge needs the predictions to be set but GG metric doesn't. So, we could either adapt LLMAsJudge or adapt GG.

[bnayahu]

@martinscooper That's true. Perhaps a more delicate way would be to add a flag field to the class (e.g., use_generated_prediction) that could be set in the specification of the metric in the card, then be checked in verify_granite_guardian_config and process_input_fields to either only use task_data or use the prediction field for the assistant field?