Dual-Modality Co-Learning for Unveiling Deepfake in Spatio-Temporal Space