PT: add uniform likelihood bucket batching

[NeoLegends]

This is a first attempt at automatically optimizing the bucket limits during training. After every subepoch the limits are adjusted to make every bucket catch a roughly equal number of segments.

I am debating on two more things:

1. Keeping only a fixed number of seq lens that have passed through the batching (e.g. 10k + upper + lower value) to have an upper bound on the memory consumption. Whenever the list is extended by another batch and crosses said limit, we re-sample randomly from that list to keep it at a fixed size. As long as the number of seq lens we keep is large enough this shouldn't have a big influence on the accuracy of the results.
2. Keeping seq len statistics even across subepochs to make the statistics better.

[albertz]

When I mention that I would try to optimize this, I was more thinking about writing a dedicated script just to do that.

But this here could also be interesting.

I'm not sure thought that I would put this into RETURNN yet, while you are still experimenting with it. You can rather put this somewhere into your i6_experiments.users. and then just use it from there. (I think we might need to extend the class resolution slightly, if "." in cls_name: ...; see similar code in rf.build_from_dict or get_optimizer_class or other such functions.)