Adaptive Cross-Modal Prototypes for Cross-Domain Visual-Language Retrieval