Faver: Boosting LLM-based RTL Generation with Function Abstracted Verifiable Middleware