Distributed learning in non-convex environments – Part II: Polynomial escape from saddle points