Abstraction and Meta-Reinforcement Learning