fix(provider): preserve Mistral reasoning history

[rekram1-node]

Summary

• Upgrade @ai-sdk/mistral to the latest AI SDK 6 release, 3.0.51
• Preserve Mistral-native thinking and text chunks instead of concatenating reasoning and visible output
• Retain and replay complete native thinking metadata, including signatures, closed state, tool references, and source references
• Preserve the existing string request shape for assistant messages that contain no reasoning
• Add provider round-trip, streaming metadata, and OpenCode LLM request-level coverage

Mistral recommends replaying the complete structured assistant message, including ThinkChunk, in multi-turn reasoning conversations: https://docs.mistral.ai/studio-api/conversations/reasoning

Follow-up to #38448.

Test plan

• bun install --frozen-lockfile
• bun test test/provider-mistral.test.ts in packages/core
• bun test test/session/llm.test.ts in packages/opencode
• bun typecheck in packages/core
• bun typecheck in packages/opencode