Improved deep learning of chaotic dynamical systems with multistep penalty losses