Proposal: simplifying structured prediction result type hinting

[ncoghlan]

The single most complex part of the current SDK type hinting is attempting to map from prediction API inputs to indicate whether or not the parsed field on the prediction result is expected to contain a plain string (unstructured) or a parsed string-keyed dictionary (structured).

This mapping is already inaccurate, since it only works off the response_format parameter input. If the response format is passed in via the config's structured field, the parsed field may still end up containing a string-keyed dictionary, but will type hint as containing a string (since response_format will be None).

Adding GBNF grammar response format support (#60) makes this problem worse, since passing in a GBNF grammar via the "response_format" field will type hint as producing a structured result, but it probably won't be a JSON result, and hence parsed will probably contain a plain string, not a string-keyed dictionary.

All of the type hinting complexity related to this can be eliminated by hinting parsed as a union type, rather than as a generic type variable.

[ncoghlan]

#61 implements this simplification (with a net removal of 330 lines of code)