Uniformly Stable Algorithms for Adversarial Training and Beyond