Towards Scalable Web Accessibility Audit with MLLMs as Copilots