Evaluating and Advancing Multimodal Large Language Models in Ability Lens