Improved Generalization Bounds for Adversarially Robust Learning