VO$Q$L: Towards Optimal Regret in Model-free RL with Nonlinear Function Approximation