Polynomial escape-time from saddle points in distributed non-convex optimization