A Multi-To-One Interview Paradigm for Efficient MLLM Evaluation