Training stochastic stabilized supralinear networks by dynamics-neutral growth