Near-Optimal Sample Complexity in Reward-Free Kernel-Based Reinforcement Learning