On Reward-Free RL with Kernel and Neural Function Approximations: Single-Agent MDP and Markov Game