[Feature request] track llm used to generate each program

[nhuet]

Currently, we do not know which program has been generated by which LLM when using a LLMEnsemble.
It would be nice to track it in metadata to compare LLM performances.

NB: When sampled we can read the selected model in log but as iterations are made in parallel we cannot relate it to the iteration number.

[HenriqueAssumpcao]

This is actually quite useful. Will try to do this by the end of this week and link a PR to this issue.