Capture reasoning_content from OpenAI-compatible Chat Completions

[CodeWithMoin]

Refs #1689

DeepSeek, vLLM and llama.cpp (with --reasoning-format deepseek) return reasoning as reasoning_content, on choices[].message or on each streamed choices[].delta. Chat and AsyncChat ignored it, so it never showed on stderr or made it into llm logs.

Now both emit it as StreamEvent(type="reasoning") before the text, streaming and non-streaming, so it goes through the same path as Responses API reasoning: dim on stderr, hidden with -R, stored as a ReasoningPart, shown under ## Reasoning in llm logs and as reasoning in --json.

One related tweak: DeepSeek also reports completion_tokens_details.reasoning_tokens, which used to add the empty redacted=True marker. That marker is now skipped when real reasoning text was emitted, otherwise the response ends up with two reasoning parts. The count-only case (OpenAI o-series) behaves as before.

Reasoning isn't sent back on later turns of a conversation, since _append_llm_message still skips ReasoningPart for Chat. That matches DeepSeek's guidance not to return reasoning_content.

Testing

New TestReasoningContent in tests/test_openai_messages.py: streamed, non-streamed, async streamed, and streamed with reasoning_tokens in usage. All four fail on main.

python -m pytest        # 1128 passed
black --check . / ruff check . / mypy llm   # clean
cog --check docs        # clean

Also ran the CLI against a local fake server streaming reasoning_content, configured through extra-openai-models.yaml like in the issue:

$ llm -m qwen38 'what is 2+2'
Two plus two is four.      <- stderr, dim
4

$ llm logs -n 1
## Reasoning
Two plus two is four.
## Response
4

$ llm logs -n 1 --json   ->   "reasoning": "Two plus two is four."

On main the same run logs "reasoning": null. With --no-stream the reasoning is logged but not printed, same as every other model, since that path only prints response.text().