eval: set USER_ERROR blame when AOAI run fails with UserError code

[Copilot]

When an AOAI evaluation run ends in a non-completed status, the exception was always raised with blame=UNKNOWN. If the run's error.code is "UserError", the blame should reflect that the failure is attributable to the user, not the system.

Changes

• _evaluate_aoai.py — _get_single_run_results: safely retrieve run_results.error.code via nested getattr and assign ErrorBlame.USER_ERROR when the code matches "usererror" (case-insensitive); falls back to UNKNOWN for any other value, missing attribute, or None.

error_code = getattr(getattr(run_results, "error", None), "code", None)
blame = ErrorBlame.USER_ERROR if isinstance(error_code, str) and error_code.lower() == "usererror" else ErrorBlame.UNKNOWN

• test_evaluate_aoai.py — TestGetSingleRunResultsBlame: five new unit tests covering:
  • "UserError" / "usererror" / "USERERROR" / "uSeReRrOr" → USER_ERROR (parametrized)
  • Non-matching code (e.g. "SystemError") → UNKNOWN
  • Missing error attribute → UNKNOWN
  • error present but no code attribute → UNKNOWN
  • error=None → UNKNOWN