Hamiltonian Deep Neural Networks Guaranteeing Nonvanishing Gradients by Design