Resurrecting the sigmoid in deep learning through dynamical isometry: theory and practice