Benchmarking Adversarial Robustness of Compressed Deep Learning Models