Episodic Control through Meta-Reinforcement Learning