Regret Bounds for Information-Directed Reinforcement Learning