League of LLMs: A Benchmark-Free Paradigm for Mutual Evaluation of Large Language Models