Adaptive Discretization-based Non-Episodic Reinforcement Learning in Metric Spaces