Provable Reduction in Communication Rounds for Non-Smooth Convex Federated Learning