Enable usage of specific GPU-instances for segmentation

[maxschloegel]

As discussed today, I would need to specify the GPU-instance I want to use for segmentation (or potentially other uses).

For now it should be enough to get this working for torch (not yet tensorflow)

To-Do:

-[ ] Understand where GPUs are used
-[ ] Implement cuda-device as parameter for relevant functions ('cuda:0', 'cuda:1', ...)
-[ ] Think of good default values, if cuda-device is not specified
-[ ] Write tests that check if correct cuda is defined and used
-[ ] Update Changelog

[maxschloegel]

@paulmueller , you can assign me.

[paulmueller]

A few additional notes from my side:

• By design of dcnum, your torch GPU segmenter will have to make sure that it is running on a GPU, or else raise an exception. There should not be a fallback-to-CPU, instead a cpu-only version should be implemented (if that is needed).
• I don't know whether this gpu_index=0 should be a general option for all GPU segmenters or not.
• Ideally, you would not have to edit any code specifically in dcnum for now, because you could just subclass from GPUSegmenter and define everything you need (including the gpu_index=0 and torch-related part) in that class.
• But there are probably bugs or cases I have not thought about in the current implementation, so there might have to be modifications of dcnum.
• GitHub Actions and GPU: probably need to have special runners or hardware defined in the .yaml files.
• The gpu_index argument should not be part of the pipeline ID, so it should not be one of the keywords-only argument...