Learning Structured Components: Towards Modular and Interpretable Multivariate Time Series Forecasting