When CNNs Outperform Transformers and Mambas: Revisiting Deep Architectures for Dental Caries Segmentation