Toward Structural Multimodal Representations: Specialization, Selection, and Sparsification via Mixture-of-Experts