Fix evaluation aggregation logic for not applicable results (Azure#42888)

* Replace not applicable results in evaluator outputs to aggregate metrics

* Update aggregation description

---------

Co-authored-by: Salma Elshafey <[email protected]>

Author: salma-elshafey

sdk/evaluation/azure-ai-evaluation/azure/ai/evaluation/_evaluate/_evaluate.py

@@ -18,6 +18,7 @@
 
 from azure.ai.evaluation._common.math import list_mean_nan_safe, apply_transform_nan_safe
 from azure.ai.evaluation._common.utils import validate_azure_ai_project, is_onedp_project
+from azure.ai.evaluation._evaluators._common._base_eval import EvaluatorBase
 from azure.ai.evaluation._exceptions import ErrorBlame, ErrorCategory, ErrorTarget, EvaluationException
 
 from azure.ai.evaluation._aoai.aoai_grader import AzureOpenAIGrader
@@ -317,6 +318,9 @@ def _aggregate_metrics(df: pd.DataFrame, evaluators: Dict[str, Callable]) -> Dic
     # For rest of metrics, we will calculate mean
     df.drop(columns=handled_columns, inplace=True)
 
+    # Convert "not applicable" strings to None to allow proper numeric aggregation
+    df = df.replace(EvaluatorBase._NOT_APPLICABLE_RESULT, None)
+
     # NOTE: nan/None values don't count as as booleans, so boolean columns with
     # nan/None values won't have a mean produced from them.
     # This is different from label-based known evaluators, which have special handling.