Efficient Restarts in Non-Stationary Model-Free Reinforcement Learning