DualFocusNet: Dual-Stream Spatio-Temporal Transformer for Generalizable Deepfake Detection | Synapse