Beyond Exponential Graph: Communication-Efficient Topologies for Decentralized Learning via Finite-time Convergence