Asymptotically Optimal Information-Directed Sampling