CI: Windows GPU runners do not stop on error

[leofang]

In this CI run we hit a bizarre NVRTC not found error at test time. However, it should have been properly installed prior to test execution. Turns out that Powershell decides to swallow any pip install failures (this happens because of #482), so we did not install the dependencies (including NVRTC) successfully:
https://github.com/NVIDIA/cuda-python/actions/runs/13623016730/job/38075976144?pr=423#step:18:39

It looks like we hit a known runner issue, which was closed without a proper fix: actions/runner-images#6668 (the recommendation there was to switch to the bash shell; I'd love to do this too as it'd allow us to not maintain 2 versions of workflows, however it is not possible for GH-hosted Windows GPU runners)

[leofang]

Not sure if setting this in the beginning of a workflow would help:

# Stop the script when a cmdlet or a native command fails
$ErrorActionPreference = 'Stop'
$PSNativeCommandUseErrorActionPreference = $true

• https://stackoverflow.com/a/9949105/2344149
• https://www.meziantou.net/stop-the-script-when-an-error-occurs-in-powershell.htm
• Set $PSNativeCommandUseErrorActionPreference = $true in a workflow that uses pwsh  kachick/dotfiles#806