feat: carry assistant reasoning on the generation output too

[milanagm]

Summary

• Follow-up to feat: record the conversation history as each generation's input #33 (LFE-15986): reasoning was already carried in a generation's history, but the message_end output was still built from extractText + extractToolCalls only, so a type: "thinking" part never reached the output of the call that produced it.
• Attaches extractThinking()'s result to the output as thinking, matching the shape the history side already uses.
• Switches extractToolCalls to the nested { id, type: "function", function: { name } } shape Langfuse renders (mirrors the Claude plugin's #77); arguments stay off the generation output since the tool span already carries them.

Linear: https://linear.app/clickhouse/issue/LFE-15989/featpi-plugin-carry-assistant-reasoning-on-the-generation-output-too
Fixes LFE-15989

Verification

• 98/98 tests + typecheck clean; no new test cases added — the new assertions were folded into the existing attaches the assistant reasoning to the step that produced it integration test (which already spawns pi), instead of adding a second pi spawn.
• Confirmed against real Langfuse traces (mock provider, no model spend), read back via /api/public/v2/observations:
  • before: https://cloud.langfuse.com/project/cmqqw1yao001iad0dohmvzben/traces/af32e1e9c81be833789f043fb2c19c04 — the reasoning call's output is role,content only
  • after: https://cloud.langfuse.com/project/cmqqw1yao001iad0dohmvzben/traces/b3e0dc74e728c5b3438cb62e9a34e07d — output gains thinking: [{"type":"thinking","content":"..."}]; tool calls are nested
• Neither trace exports thinkingSignature or the raw reasoning_content field.

Not in this change

• opencode's tool_calls are still flat ({id, name, arguments}) despite using the same ChatML thinking shape — worth its own issue, out of scope here.