Hamiltonian Monte Carlo on ReLU Neural Networks is Inefficient