Weakly Supervised Hashing with Reconstructive Cross-modal Attention