Learning Robust Algorithms for Online Allocation Problems Using Adversarial Training