[WIP] Guidance integration for more reliable optimization

[mertyg]

We had a set of issues (e.g., #15 ) where the optimizer fails to follow instructions when sending the new state of a variable. This happens in particular when we have long context / variables with very long states and less capable models fail to follow the optimizer instructions.

For this purpose, we'll try integrating Guidance to directly constrain the generations.

• Guidance Engine implementation (uses huggingface transformers)
• GuidedTextualGradientDescent implementation
• An example notebook / snippet to show how to use the two
• Perhaps more broadly integrating guidance for other types of structured generation
• Better error handling for when guidance fails