Deciphering the Role of Representation Disentanglement: Investigating Compositional Generalization in CLIP Models | Synapse