RL²: Fast reinforcement learning via slow reinforcement learning