Interpreting Neural Networks through the Polytope Lens