fix(chat): LMStudio fallback + remember base URL; improve logging and error surfacing

[HenkDz]

Summary

• Default to LMStudio base URL when headers are missing, and remember the last working base URL from the models list for chat.
• Surface non-retryable streaming errors (e.g., 401) to the client with clear, actionable messages.
• Make Windows logging UTF-8 safe and provide a global logger instance.
• Add diagnostics: log model-list HTTP errors.

Why

• LMStudio users could list models but chat failed when headers weren’t forwarded; this enables out-of-the-box local usage.
• Streaming errors were opaque; now users see explicit reasons and next steps.
• Windows consoles/files previously hit cp1252 Unicode errors; resolved with UTF-8-safe handlers.

Changes

• backend/controller/conversation_api.py: LMStudio/default base URL fallback; early auth messaging if neither key nor LMStudio URL is present.
• backend/utils/globals.py: add LAST_OPENAI_BASE_URL with getters/setters.
• backend/controller/llm_api.py: remember base URL after successful models list; log failures.
• backend/service/mcp_client.py: propagate non-retryable error text to the stream.
• backend/utils/logger.py: UTF-8-safe logging; export global log.

Testing

• Model list works with LMStudio set in settings.
• Chat runs without an OpenAI key when LMStudio base URL is present or remembered.
• Unicode logging no longer crashes on Windows.

Notes

• No schema changes. UI still passes headers when configured; backend has sensible defaults for local LMStudio.

[ai-business-hql]

I'll test it as soon as possible.

[ai-business-hql]

You changed a lot, I'm afraid that my function be affected, so I'll test fully, and need to take some time.