Non-Stationary Reinforcement Learning: The Blessing of (More) Optimism