Text this: Adversarial classification via distributional robustness with Wasserstein ambiguity