fix: resolve AI chat hang caused by main-actor-bound streaming (#735)

[datlechin]

Summary

• Move streaming loop, schema fetching, and conversation loading off @MainActor using Task.detached with explicit MainActor.run for state mutations — fixes the primary hang
• Cap concurrent schema fetch queries to 4 (sliding-window TaskGroup) instead of firing all 20 at once against a serial driver connection
• Consolidate JSON parsing in AnthropicProvider (3→1 parse per SSE line) and OpenAICompatibleProvider (2→1 parse per SSE line)
• Replace per-token scroll onChange with stream-end scroll to eliminate 50-100+ layout passes/sec during streaming
• Wrap AIProviderFactory cache with OSAllocatedUnfairLock for thread safety
• Remove @MainActor from AISchemaContext.buildSystemPrompt — pass editorLanguage/queryLanguageName as params instead
• Fix startStreaming not cancelling existing stream before starting a new one
• Fix fetchSchemaContext race when tables change rapidly (store + cancel task)
• Fix cancelStream using messages.last instead of streamingAssistantID for placeholder removal
• Handle Anthropic in-stream error events (e.g., overloaded_error) instead of silently ignoring them

Closes #735

Test plan

• Connect to PostgreSQL (Docker), open AI chat, send a message — UI stays responsive during streaming
• Test with Ollama (fast local streaming) for highest token rate — no spinning cursor
• Open AI panel with 20+ tables — schema fetch doesn't stall the UI
• Start streaming, tap Stop mid-stream — clean cancellation, partial content preserved
• Send a new message while previous stream is still active — old stream cancelled, new one starts cleanly
• Switch databases rapidly — schema context updates correctly without stale data
• App launch with saved conversations — conversations load without blocking startup
• Test all providers: Anthropic, OpenAI, Ollama, Gemini — streaming works correctly