CAARL: In-Context Learning for Interpretable Co-Evolving Time Series Forecasting