Accelerating Decentralized Deep Training with Sparse and Effective Topologies