Leveraging Self-Supervised EEG Foundation Models for Data-Efficient and Clinically Robust Neurological Diagnostics