feat(http_provider): enhance K2-Think model support

[shauryr]

Added reasoning_effort parameter and handling reasoning_content in responses

[lukemilby]

Tech document referencing the need for https://platform.moonshot.ai/docs/guide/use-kimi-k2-thinking-model#accessing-the-reasoning-content

@shauryr have you test this with other models that are used by the HTTP Provider?

[shauryr]

Thanks for reviewing! Quick clarification; LLM360/K2-Think-V2 https://huggingface.co/LLM360/K2-Think-V2 what I tested and Moonshot Kimi K2 are different models. The changes are generic and support any reasoning model.

Here are the specific change I made which should be generic (all models):

• Add reasoning_content field parsing
• Fallback to reasoning_content if content is empty
• Strip special tokens

K2-Think specific:

• reasoning_effort: "high" parameter

Also the changes are backward compatible

• Standard models: content has data so it uses content
• Reasoning models: content is empty so it uses reasoning_content

I have done the testing for - K2-Think-V2 via vLLM (basic queries + tool calling) as well as multi-step reasoning with tools. Also this should Work With DeepSeek R1, QwQ, and other models using reasoning_content

Happy to test with specific models if you have recommendations!