Rethinking Channel Dependence for Multivariate Time Series Forecasting: Learning from Leading Indicators