Indicators of Attack Failure: Debugging and Improving Optimization of Adversarial Examples