CATS: Mitigating Correlation Shift for Multivariate Time Series Classification