Bayes-Adaptive Deep Reinforcement Learning via Meta-Learning