Neural ODEs as Feedback Policies for Nonlinear Optimal Control