From Initial Data to Boundary Layers: Neural Networks for Nonlinear Hyperbolic Conservation Laws