Provably Efficient Reinforcement Learning with Kernel and Neural Function Approximations