Improving LLM-based Document-level Machine Translation with Multi-Knowledge Fusion