Gravar-mail: Learning Optimal Distributionally Robust Individualized Treatment Rules