Dataset-Driven Channel Masks in Transformers for Multivariate Time Series