Stochastic maximum principle for equations with delay: the non-convex case