From Generative to Episodic: Sample-Efficient Replicable Reinforcement Learning