Adversarially Robust Learning for Security-Constrained Optimal Power Flow