Response Hallucination: decouple LLMs for sample generation and consistency checking

[FelipeAdachi]

Currently, the response_hallucination feature uses the same LLM for both additional samples generation and consistency checking. We should support being able to define different LLMs for each process.