Harnessing RLHF for Robust Unanswerability Recognition and Trustworthy Response Generation in LLMs