Cross-feature Contrastive Loss for Decentralized Deep Learning on Heterogeneous Data