On the Statistical Efficiency of Reward-Free Exploration in Non-Linear RL