Landmark-Guided Gaze Estimation via Conditional Keypoint Generation and Cross-Attention Fusion | Synapse