Interpretable LLM-based Table Question Answering