Towards More Efficient Stochastic Decentralized Learning: Faster Convergence and Sparse Communication