Create a new subclass of Domain to support open-ended tuning

The new subclass supports sampling a variation of the current text, based on feedback. This new subclass of Domain would be useful for LLM-based optimization. For example, to allow open-ended tuning of prompt using LLM, including #1063.

The new subclass can use an LLM-based sampler. In a search algorithm like CFO, we can pass textual feedback from results to the sampler and use LLM to sample a new variation.