Differentially Private Decentralized Learning with Random Walks