A Kernel-Based Approach to Non-Stationary Reinforcement Learning in Metric Spaces