Practical Low-Rank Communication Compression in Decentralized Deep Learning