COEF-VQ: Cost-Efficient Video Quality Understanding through a Cascaded Multimodal LLM Framework