FIX: Use underlying_model_name for evaluation identifier and add param_fallbacks

[jsong468]

There existed a bug where ScorerEvaluationIdentifier used model_name instead of underlying_model_name as originally intended. Thus, the eval hash in our JSONL scorer metrics registries is computed incorrectly using model_name (which is often a deployment name).

This PR addresses the bug by using underlying_model_name in ScorerEvaluationIdentifier while also introducing param_fallbacks field to ChildEvalRule to use model_name as a fallback for underlying_model_name when it is None or "" (for cases where the user doesn't provide a separate underlying_model_name to the LLM prompt target used in a scorer since it now needs to be passed in explicitly and not automatically from environment variables as a result of #1590)

Description

• Eval hash now uses underlying_model_name instead of model_name to group scorer evaluations by the actual model (e.g., "gpt-4o") rather than the deployment name (e.g., "my-azure-deploy")
• Added param_fallbacks field to ChildEvalRule to handle cases where underlying_model_name is empty — falls back to model_name
  • This is necessary because users don't often explicitly set underlying_model, so without fallback, all targets would hash to the same empty string
• Fallback logic uses explicit is None / == "" checks instead of truthiness to avoid incorrectly triggering on valid values like 0 or False
• Updated ScorerEvaluationIdentifier and AtomicAttackEvaluationIdentifier rules to use underlying_model_name with model_name fallback
• Change is contained entirely in the eval layer; there is no impact on full ComponentIdentifier

Tests and Documentation

• Added TestParamFallbacks class in test_evaluation_identifier.py covering: primary param used when present, fallback on empty string, fallback on missing key, and no fallback when param_fallbacks is None
• Added TestScorerEvalHashFallback class in test_scorer_evaluation_identifier.py covering: underlying model used when present, fallback to model_name when empty/missing, and different models produce different hashes
• Updated existing tests in test_scorer_evaluation_identifier.py and test_atomic_attack_identifier.py to reflect the new underlying_model_name + param_fallbacks rules

TODO

• re-run evaluate_scorers.py script and re-populate all JSONL registries since every eval hash will be different.

[adrian-gavrila]

Other than the one comment this looks great