Improve custom provider compatibility and agent concurrency

[cube1345]

Summary

• Improve the direct OpenAI-compatible custom provider with stable session affinity, configurable extra headers, request timeouts, clearer error classification, and compatibility retries for unsupported reasoning_effort / tool_choice parameters.
• Replace the agent loop's global processing lock with per-session locks so different sessions can run concurrently while preserving message ordering within the same session.
• Add structured JSON tool errors with trace IDs for easier debugging.
• Keep CLI status/startup offline-friendly by forcing LiteLLM to use its local model cost map.

Tests

• env -u ALL_PROXY -u HTTP_PROXY -u HTTPS_PROXY -u all_proxy -u http_proxy -u https_proxy PYTEST_DISABLE_PLUGIN_AUTOLOAD=1 UV_CACHE_DIR=/tmp/uv-cache uv run pytest -p pytest_asyncio.plugin tests/test_provider_factory.py tests/test_task_cancel.py tests/test_tool_validation.py tests/test_config_migration.py