AutoJudger: An Agent-Driven Framework for Efficient Benchmarking of MLLMs