Sync snowfall with Lhotse's dataset/sampler refactoring PR

[pzelasko]

Reflects the changes from lhotse-speech/lhotse#194

This new version allows working with distributed training out-of-the-box. Once it is merged, I will rebase #71 on that and finish multi-GPU training support.

One thing I am concerned about is the number of files with duplicate code. It becomes quite labor-intensive to propagate these kinds of changes, and I am not sure if I'm doing everything OK as I can't run all of these (🙏🏻 static checker tool helps 🙏🏻). It could be good to slowly converge to some common code structure that can be shared and imported (e.g. a Trainer class or something like that)?

[danpovey]

How do you run the static checker tool?
I'm OK with leaving snowfall fairly unstructured. As far as I am concerned it is just a repo where we can iterate rapidly on experiments, and not a serious thing that we are "releasing"; we can do a cleaned-up repo later once we have an idea what kinds of models we'll be using.

[pzelasko]

I use the one that is integrated into PyCharm and IMO it is the best tool available. I believe the next best thing is mypy, and it has integrations with multiple editors/IDEs. Years ago when these things were not available I used flake8 it says it's "style enforcement" but it also has some static checks (and a rich plugin ecosystem I think). In case you are using VSCode, perhaps pylance is interesting (it's exclusive to VSCode though).

[pzelasko]

OK this works - merging.