Information-Directed Exploration for Deep Reinforcement Learning