FaVChat: Unlocking Fine-Grained Facial Video Understanding with Multimodal Large Language Models