Mitigating Overfitting in Medical Imaging: Self-Supervised Pretraining vs. ImageNet Transfer Learning for Dermatological Diagnosis