Cross-modal attention preservation with self-contrastive learning for composed query-based image retrieval