On the Algorithmic Stability of Adversarial Training