Second-order guarantees of stochastic gradient descent in non-convex optimization,