[Evaluation] Normalize evaluator validation errors to EvaluationException with USER_ERROR blame

[m7md7sien]

Summary

Normalizes evaluation/validation error handling in azure-ai-evaluation so that user input and configuration errors are consistently raised as EvaluationException with blame=ErrorBlame.USER_ERROR (plus an appropriate category and target).

Previously several evaluators raised bare ValueError/TypeError for input/threshold validation, and a few existing EvaluationException raises did not set blame, so they defaulted to Unknown/InternalError even though they were caused by user input.

Changes

Raw ValueError/TypeError → EvaluationException(USER_ERROR)

• ContentSafetyEvaluator — threshold type check
• QAEvaluator — threshold type check
• RougeScoreEvaluator — threshold type check
• DocumentRetrievalEvaluator — ground-truth label and input-record validation
• Task navigation efficiency evaluator — matching_mode and ground_truth validation

Existing EvaluationException missing USER_ERROR

• Evaluator base (_base_eval.py) — conversation message mismatch, malformed tool-call parsing, and threshold-not-a-number checks now set blame=USER_ERROR (one category=UNKNOWN corrected to INVALID_VALUE)

Supporting

• Added QA_EVALUATOR, ROUGE_EVALUATOR, and DOCUMENT_RETRIEVAL_EVALUATOR members to ErrorTarget
• Updated the task navigation test to expect EvaluationException for an invalid matching_mode
• CHANGELOG entry under 1.17.1 (Unreleased)

Intentionally left unchanged (not user errors)

• "Evaluator returned invalid output" / "Invalid score value" across the prompty and tool evaluators remain SYSTEM_ERROR (malformed LLM output, not user input).
• Internal/defensive checks (_conversation_aggregators.py UNKNOWN, _base_rai_svc_eval.py "Not implemented") are unchanged.

Validation

• All affected unit tests pass (document retrieval, task navigation, threshold behavior, common validators, built-in & agent evaluators).
• black (pinned 24.4.0, repo config) passes on all modified files.