Parametric Complexity Bounds for Approximating PDEs with Neural Networks