fix: vertex files, thinking export, completions port

[Im-Beast]

• Vertex AI adapters now send file history as inline data instead of the Files API (which Vertex doesn't support)
  • fixes prod file-attached chats dying with "only supported by the Gemini Developer API"
• Fixed inverted catch in ensureFileUploaded — the Files-API storage cleanup ran for every error except the quota one it was written for. Non-quota errors now rethrow.
• Exported getThinkingConfig (+ GoogleModels, SupportedThinkingLevel) from ./google-genai so wrappers building their own Vertex adapters can map thinking levels
• Ported the openai_completions fixes from our pnpm patch: DeepSeek reasoning_content streaming, and batching a turn's tool calls into one assistant message with reasoning re-attached