support parsing/validating openai responses

[tarasglek]

we like to use chatcraft for mermaid chart creation but there are often subtle syntax errors.

should make it easy to send those errors with maximum context back to openai to get a revised ver.

alternatively we can support streaming parsers which abort on invalid output and send rhat to openai automatically

similarly it would be cool to do streaming transforms of json. This opens an option to require valid and send back error json same as above.

Also gives an option to render them as yaml as models are better at generating json for now

this opens path to using grammar restrictions in llama.cpp on openai output