Provably Efficient RL with Kernel and Neural Function Approximations