Contrastive Speaker-Aware Learning for Multi-party Dialogue Generation with LLMs