On Generalization of Decentralized Learning with Separable Data