Breaking Neural Network Scaling Laws with Modularity