fix: modifies openai request logic for reasoning models (#4221)

[myaple]

This is a first take at fixing the issue described in #4221 where the logic used in the litellm provider to create the requests is the same as the openai format, which has significant bias towards openai models. I've added a GOOSE-level configuration, GOOSE_REASONING_EFFORT, that forces goose to use the reasoning level provided, or it falls back to the previous parsing style for the model name. The name-based parsing uses more specific schemes, calling out o{1..4} instead of just "o" as the starting char.

I also split the logic for "is this a reasoning model" and "is this an ox model" as those have different implications for the built request. I'm open to the idea of a better way to distinguish whether or not a model is reasoning for models other than openai-specific ones, as I'm not really sure - user config parameter, enumerate the most common reasoning models and try to parse them, not sure what else.