Generalization Bounds for Neural Networks via Approximate Description Length