Practical Low-Rank Gradient Compression in Decentralized Deep Learning