Incorporating unlabeled data into Wasserstein distributionally-robust learning