Improving Robustness of Reinforcement Learning for Power System Control with Adversarial Training