Efficient variance-reduced learning over multi-agent networks