Further Optimal Regret Bounds for Thompson Sampling