Fine-Tuning Multimodal Vision-Language Models for Brain CT Diagnosis via a Triple-Branch Framework