Multiple-Choice Questions are Efficient and Robust LLM Evaluators