AFL-Net: Integrating Audio, Facial, and Lip Modalities with Cross-Attention for Robust Speaker Diarization in the Wild.