Understanding when Dynamics-Invariant Data Augmentations Benefit Model-free Reinforcement Learning Updates.