Quantifying Modality Contributions via Disentangling Multimodal Representations