Domain-Robust Visual Imitation Learning with Mutual Information Constraints