Learning Adversarially Robust Representations via Worst-Case Mutual Information Maximization