Run gradient descent on a nonconvex scalar loss and report the optimization path and final stationary behavior.
