Adversarial Robustness with Semi-Infinite Constrained Learning