The Curious Price of Distributional Robustness in Reinforcement Learning with a Generative Model