Consider saving raw predictions in benchmarks

[PSSF23]

Saving raw predictions of classification tasks helps improving evaluation metrics (changing/adding/deleting/...). As the test sets are randomly generated, test labels need to be saved at the same time.