Learning Rate-Free Reinforcement Learning: A Case for Model Selection with Non-Stationary Objectives