Perturbation-mitigated USV Navigation with Distributionally Robust Reinforcement Learning