Add support for OpenAI functions

[rlouf]

We would like to integrate OpenAI functions to make it easy to compare guided generation on OSS model with the feature offered by OpenAI. Here are the design requirement:

• Usage should be very similar to guided generation, and users should be able to change a single line to switch between OpenAI and OSS models. Since OpenAI's function API does not necessarily return valid JSON I am not sure about dispatching outlines.text.generate.json to different functions depending on the model type. This however makes the model switch even easier;
• When inferring the output format from the function's docstring we should extract the docstring and add it as the model's description since OpenAI uses this information in their prompt;
• Output format should be the same as with the current outlines.text.generate.json interface: Pydantic model instance when the input format is a Pydantic model, dictionary when it is a function or a JSON schema.

[A-F-V]

@rlouf I am happy to pick this one up if no one else is working on it :)

[rlouf]

Would be awesome to add an interface to function calling that uses generate.json 😁 Go for it!

[A-F-V]

I would be up for trying it but I can't setup the repo locally on my mac as per this thread: https://discordapp.com/channels/1182316225284554793/1252592587387310080

[A-F-V]

c.r. #989