Improving Plasticity in Non-stationary Reinforcement Learning with Evidential Proximal Policy Optimization