CUDA memory not freed after OOM error

[qin-yu]

1. Either it should be reported in auto patch size finder
2. Or it should be freed before the next prediction

[qin-yu]

Should be able to add tests for this and check with GitHub Actions now:

• Run your ML workloads on GitHub Actions with GPU runners
• https://github.com/github/roadmap/issues/961#issue-2262146959