Interpretability in Parameter Space: Minimizing Mechanistic Description Length with Attribution-based Parameter Decomposition