Redact model config credentials in saved and returned results

[yangbaechu]

Summary

This PR redacts credential fields from model_config before they are exposed through evaluation result artifacts.

Specifically, it masks:

• api_key (LiteLLM)
• inference_server_auth (TGI)

The redaction is now applied to both:

• saved results via EvaluationTracker.results
• returned results via pipeline.get_results() / generate_final_dict()

Problem

EvaluationTracker currently serializes model_config into config_general without redacting credential fields. As a result, values like api_key and inference_server_auth can be persisted in saved result artifacts and exposed through pipeline.get_results().

Changes

• redact api_key and inference_server_auth from serialized model_config
• apply the same redaction policy to both saved results and returned results
• add regression tests to verify that:
  • LiteLLM api_key is redacted in results
  • TGI inference_server_auth is redacted in results
  • LiteLLM api_key is redacted in pipeline.get_results()
  • TGI inference_server_auth is redacted in pipeline.get_results()