Physics-Informed Singular-Value Learning for Cross-Covariances Forecasting in Financial Markets