Understanding Measures of Uncertainty for Adversarial Example Detection