From Single to Multi: How LLMs Hallucinate in Multi-Document Summarization