Describir: Learning Self-distilled Features for Facial Deepfake Detection Using Visual Foundation Models: General Results and Demographic Analysis