The strict saddle property states that every saddle point has at least one direction of strict negative curvature. Why is this property useful for gradient descent analysis?