fix: encode thinking content parts as reasoning_content for OpenAI-compatible providers

[spfoos]

When assistant messages contain ContentPart.thinking(...) parts (e.g. from Moonshot reasoning models), the OpenAI encoder was stripping them via encode_openai_content_part returning nil for :thinking parts.

This caused multi-turn conversations with Moonshot to fail with:

thinking is enabled but reasoning_content is missing in assistant tool call message

Changes:

• Add extract_reasoning_content/1 to pull thinking parts out of message content
• In encode_openai_message, extract reasoning_content before encoding content, then add it as a top-level reasoning_content field on the encoded message
• Remove thinking parts from the content list so they don't get encoded as nil

This matches the OpenAI/Moonshot API contract where reasoning_content lives at the message level, not inside the content array.

[mikehostetler]

Thanks for the focused fix here. I marked this needs_work because the encoder change needs a small regression test proving assistant thinking parts are emitted as top-level reasoning_content and removed from the content array. It would also help to either scope that field to the OpenAI-compatible providers/models that accept it, or add a test/note showing the shared encoder behavior is intentional across those providers. Appreciate the contribution; this looks close once that coverage and scope are tightened up.