For LLM, RAG and other model nodes return whole output

Currently they only return the generated text, we should return the tokens used, etc