Implement CoT no-op for reasoning models

[FireMasterK]

• Add an optional parameter to allow specifying whether a model is a reasoning model or not. If not specified, we to determine it using litellm.supports_reasoning.
• I had to upgrade to upgrade the LiteLLM library to have access to the function, as it wasn't present in the version used earlier.

Additional context: https://x.com/DSPyOSS/status/1931886536221724848

[okhat]

This is AMAZING. Thank you @FireMasterK !!

I think we'll need to think (no pun intended) about reasoning models like Qwen3 that expose their reasoning in the string... What does LiteLLM do for that. I guess it says False?

Also I guess we should still return a reasoning field, but set it to something that indicates that the reasoning is not available.

[FireMasterK]

I think we'll need to think (no pun intended) about reasoning models like Qwen3 that expose their reasoning in the string... What does LiteLLM do for that. I guess it says False?

I don't think they support telling whether a model displays its CoT or not, which is unfortunate.

It also seems like LiteLLM doesn't have enough mappings for reasoning models (eg. Qwen3 models are not there), which means we can't solely rely on LiteLLM telling if a model supports reasoning or not.
Source: https://github.com/BerriAI/litellm/blob/main/model_prices_and_context_window.json

Also I guess we should still return a reasoning field, but set it to something that indicates that the reasoning is not available.

Yes, but some model providers providers summarize the CoT, what should we do about these? I have added the parsing of reasoning_content from non-streaming responses. I'm not sure what should be done for reasoning content. Should I create a new class like dspy.streaming.ReasoningStreamListener?

[FireMasterK]

I fixed the serialization issues, so all tests now pass! Let me know what I should do about #8375 (comment).

I'm happy to create another PR or add them in this PR. Whichever you prefer.