Assessing Neural Network Robustness via Adversarial Pivotal Tuning