fix(device): isolate ONNX provider probing

[xy200303]

• I have read and agree to the CLA.

Summary

• move ONNX Runtime provider probing into a hidden child command instead of importing it in the main process
• fall back to CPU cleanly when provider probing times out, crashes, or returns invalid output
• add regression coverage for successful probes, failures, timeouts, invalid payloads, and provider caching

Verification

• pytest -q tests/test_device_manager.py
• python -c "...module._probe_onnx_providers(timeout=3)..." (confirmed the probe returns safely in this Windows environment)

Notes

• this addresses the crash pattern reported in windows 使用X-AnyLabing GPU版本时点击ai选择模型软件会闪退 #1386 when opening AI model selection on Windows GPU builds
• it also hardens startup and checks-style device queries against ONNX Runtime import crashes in the parent process