Sampling random generator

[sfalkena]

Following #2307, adding the possibility to set the seed of the random samplers by supplying a torch.Generator. As far as I could tell, this is the only place where no random generator was used yet, apart from the Kornia transforms (would we want that? They do not support a generator argument).

[adamjstewart]

Looks great! Just some minor comments on consistency and simplicity.

We should probably do the same for all datamodules, but let's do that in a separate PR.

[sfalkena]

What is the thing with codecov failing? When I check the dashboard states 100%.

[adamjstewart]

I'm trying to figure out the codecov thing now. I'm 99.9% sure it's a glitch, not your fault. Let's rerun the tests just to make sure.

[adamjstewart]

default: None

This is actually a bit tricky. For many uses of torch.rand*, None is fine. However, in some cases like data module splitting, we actually have to use some kind of generator. If we don't, we end up with overlapping train/val/test splits. So to clarify, I think the default value to use in the function signature is None, but for some cases, we need to use:

if generator is None:
    generator = torch.default_generator()

or something like that.

[adamjstewart]

Do you want to open a PR to harmonize the previous examples where we used generator?