fix(providers): echo reasoning content back for thinking-mode models

[rexlunae]

Fixes the chat error seen with DeepSeek (and other thinking-mode OpenAI-compatible providers):

Web stream error for model 'deepseek-v4-flash (adapter: OpenAI)'.
Status: 400 Bad Request
Body: {"error":{"message":"The `reasoning_content` in the thinking mode must be passed back to the API."}}

Root cause

RustyClaw captured the model's reasoning only for the live thinking display and dropped it from the turn. DeepSeek requires the reasoning_content of earlier assistant messages to be echoed back verbatim on every follow-up request, so both the tool loop and next-turn history replay hit a 400.

Changes

• ModelResponse carries the reasoning text: the streaming path accumulates ReasoningChunk events, the batch path captures ContentPart::ReasoningContent.
• The canonical assistant_tools envelope stores the reasoning; decode_assistant re-emits it as a genai ReasoningContent part, which the genai OpenAI adapter serializes back as the sibling reasoning_content field (the Anthropic adapter echoes it as a thinking block, which its multi-turn API also requires).
• New assistant_content() helper: bare text turns stay plain text; turns carrying reasoning/tool calls use the envelope.
• ThreadMessage persists the reasoning (serde-defaulted, skipped when absent — existing threads.json files load unchanged). Final assistant turns are stored with their reasoning (main chat dispatch + cron runs), and thread_history_to_chat_messages replays it, so follow-up turns and fresh sessions work.

No wire/protocol changes: ChatMessage and all frames are untouched.

---