fix(openai-formats): extract reasoning_content in non-streaming responses

convertOpenAIResponseToAnthropic never read choice.message.reasoning_content,
so DeepSeek and other reasoning-capable openai-compatible providers silently
dropped thinking content on non-streaming responses (the streaming path in
stream.ts already handled this correctly). Now emits a leading `thinking`
content block when reasoning_content is present, matching Anthropic's block
ordering. No gate needed — the field is simply absent for providers that
don't return it, so behavior for existing openai-compatible providers is
unchanged.

Addresses the reasoning-detection part of #434.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>