Bump datasets floor to >=5.0 (older versions can't parse the List feature type in this dataset's parquet metadata)

evaluation_kit/requirements.txt

@@ -2,5 +2,5 @@
 # from the Hugging Face Hub). Every other script in this kit — including
 # convert_predictions_to_evaluation_csv.py and everything under evaluation/
 # and box_grouping/ — is pure standard library, no dependencies required.
-datasets>=2.14
+datasets>=5.0
 pillow>=10.0