LaViC: Adapting Large Vision-Language Models to Visually-Aware Conversational Recommendation