Control reasoning, test for different models

[johann-petrak]

See https://docs.litellm.ai/docs/reasoning_content

We have to check how this will work with out implementation and especially also tooling implementation.

[johann-petrak]

IMPORTANT: setting some (all?) of the reasoning related parameters seems to require drop_params=True
By default we set drop_params=False which can help to detect unsupported parameters in the config file or call. However, if drop_params=False is mentioned in the config or the llms.query call, we do not change it.

For example we can now set:

• reasoning_effort=dict(effort="high")
• thinking=dict(type="enabled", budget_tokens=1024)

etc. if the model supports this. See the LiteLLM docs for details.

If return_response=True we can also get the reasoning content and/or the thinking blocks if the model created them: