A Lifted Bregman Formulation for the Inversion of Deep Neural Networks