Privacy Loss of Noisy Stochastic Gradient Descent Might Converge Even for Non-Convex Losses