fix: stabilize live error handling and quota recovery

[highkay]

Summary

• persist rate-limited quota exhaustion and refresh accounts on demand after 429 responses
• restore account windows from stored quota metadata instead of basic-tier defaults, and normalize request validation plus transport failures into client-safe API errors
• default public OpenAI/Anthropic compatible endpoints to non-streaming JSON unless stream=true is explicitly requested

Testing

• docker run --rm -v /home/admin/grok2api-origin-pr:/workspace -w /workspace ghcr.io/chenyme/grok2api:latest python -m unittest tests.test_account_refresh_rate_limit tests.test_account_on_demand_refresh_trigger tests.test_account_selector_window_reset tests.test_transport_error_wrapping tests.test_public_api_stream_defaults