Probing Multimodal Large Language Models for Global and Local Semantic Representations