Re-Bench: Evaluating ML agents against human ML experts