A lattice-based approach to the expressivity of deep ReLU neural networks