Accept JSON string inputs in TaskNavigationEfficiencyEvaluator

[Copilot]

Port of Azure/azureml-assets#4999 to the SDK repo. TaskNavigationEfficiencyEvaluator only accepted response/ground_truth as native Python lists/tuples, so callers passing serialized JSON strings (e.g. from data pipelines) failed validation.

Changes

• _evaluators/_task_navigation_efficiency/_task_navigation_efficiency.py: At the top of _do_eval, if response or ground_truth is a str, attempt json.loads and replace the value on success.
• _evaluators/_common/_validators/_task_navigation_efficiency_validator.py: Same fallback at the top of validate_eval_input (the SDK uses a separate validator class, not an inline method like the upstream evaluator).
• tests/unittests/test_task_navigation_efficiency_evaluators.py: Added test_json_stringified_valid_inputs covering the EXACT_MATCH path with JSON-stringified inputs.
• CHANGELOG.md: Bug fix entry under 1.16.7 (Unreleased).

On parse failure the original string is left in place so downstream validation surfaces the error as before.

for key in ("response", "ground_truth"):
    value = eval_input.get(key)
    if isinstance(value, str):
        try:
            eval_input[key] = json.loads(value)
        except (ValueError, TypeError):
            pass

Note: the SDK uses response/ground_truth parameter names rather than actions/expected_actions from azureml-assets; keys were adapted accordingly. No spec.yaml version bump applies here.