What changes are proposed in this pull request? The mlflow.evaluate API includes a predictions argument that selects the column from the return of predict_fn when it is a dataframe. That logic was previously implemented in _extract_output_and_other_columns as part of the default evaluator. When #13360 refactors the evaluators into multiple class, that function was omitted from the regressor evaluator. #14388 intended to fix it but simply threw an error if predict_fn returns a dataframe. This PR proposes reusing _extract_output_and_other_columns from the default evaluator to handle different format types in the predict return type.