CUDA error: CUBLAS_STATUS_NOT_INITIALIZED

**System Information (please complete the following information):**
- Model Builder Version (available in Manage Extensions dialog): Latest
- Visual Studio Version 2022

**Describe the bug**
- On which step of the process did you run into an issue: After clicking train in Text Classification with GPU
- Clear description of the problem: I receive the attached error
Dataset has one lable (multiclass, up to 3000 categories), one text field.

It there is a limit of max categories, then the error is probably valid but could be more user-friendly to help know what to do.
If I change to label with 100 or 500 categories, then I get another error, see below.

I think I once received an error with 10 labels, too, but I may be mistaken. I think the Cancel button will give an error message so maybe that got me confuder. However, it may work with 10 labels, I will update tomorrow after letting it run over the night.

[jobads-isco-29102022-partial-predicted-isco-full-level-AQ87A5.txt](https://github.com/dotnet/machinelearning-modelbuilder/files/9985315/jobads-isco-29102022-partial-predicted-isco-full-level-AQ87A5.txt)


Next time I tried I got this error module_to_device

Another error which happened during testing, it may have been related to cancelling a training

Beginne damit, das Training von Text Classification mit GPU unter Verwendung des beigefügten Datensatzes und der gemeldeten Kategorienanzahlen zu reproduzieren. Überprüfe die im Bericht und in den Screenshots gezeigten Fehler CUBLAS_STATUS_NOT_INITIALIZED, module_to_device und cancellation. Die Aufgabe ist erledigt, wenn der Fehler reproduzierbar ist und die resultierende Fehleranleitung die nächste Aktion klar erklärt.

Vom Indexierungsmodell aus dem Issue-Text verfasst.