Improving the Consistency in Cross-Lingual Cross-Modal Retrieval with 1-to-K Contrastive Learning