Build Karabo and it's dependencies against CUDA

[Lukas113]

After reading conda cuda builds, I think we should do the same (if I understood the article correctly, please correct me if I may have misunderstood something). This would have the following advantages I think:

• We wouldn't rely on a specific toolkit-version where some users might have issues if they a gpu-driver which is not up to date (even though the minimum required driver-version is only dependent on the major version of cuda-toolkit (see cuda compatibility).
• We wouldn't have to worry about fixed-toolkit which could constrain us on which package & version we intend to include

This would imply changing the conda-build recipe of karabo and cuda-enabled packages of the Feedstock as well as changing the github-workflows and the dockerfiles.

[sfiruch]

Wouldn't this mean we'd have to test against all CUDA versions?

Using the oldest-possible CUDA versions means we're compatible with most drivers. That looks like a win. This proposal would result in Karabo crashes for users with out-of-date drivers.

[Lukas113]

If we can test against CUDA (which we don't do at the moment because the CI doesn't have GPUs), I think we shouldn't have to care about the toolkit-version because conda should identify the highest compatible version according to the documentation.

[Lukas113]

That's at least my understanding about the text written in conda cuda builds I've posted in the first comment.