Benchmarking the Attribution Quality of Vision Models