Rethinking Adversarial Attacks in Reinforcement Learning from Policy Distribution Perspective