Towards generalizable object counting via dual-prompted cross-modality fusion network