Doubly Robust Augmented Transfer for Meta-Reinforcement Learning