Demystifying the Adversarial Robustness of Random Transformation Defenses