Should We Fine-Tune or RAG? Evaluating Different Techniques to Adapt LLMs for Dialogue