Adding example how to use llm ensemble estimator (for prompt squeeze)

[Eladlev]

1. Refactor: estimator doesn't modify the dataset directly.
2. Add new llm_batch estimator
3. Add the correspond config file
4. Improve the meta-prompt to focus on large errors