Towards Evaluating and Building Versatile Large Language Models for Medicine