Towards efficient compression and communication for prototype-based decentralized learning