fix(ai): defer Mistral thinking metadata until block end - #52008

Summary

• Buffer Mistral native thinking units with append-only parser state rather than copying the accumulated array per chunk.
• Keep reasoning deltas incremental and attach the complete native thinking block once at reasoning end for replay.
• Preserve metadata-only blocks, native fields, and text/thinking ordering.

Verification

• bun test test/provider/mistral-chat.test.ts test/provider/mistral.recorded.test.ts in packages/ai (23 passed)
• bun test --timeout 30000 --only-failures in packages/ai (1,607 passed, 28 skipped)
• bun typecheck in packages/ai and bun run check at the root
• Local-only 1,024-chunk stream: 1,024 incremental deltas without cumulative metadata, one final native block with all 1,024 units