Hypothetical Documents or Knowledge Leakage? Rethinking LLM-based Query Expansion