RF: weight parametrization

[albertz]

Fix #1518.

• Add rf.gradient_checkpoint_scope.
• Add rf.register_parametrization, similarly as in PyTorch.
• Add rf.weight_dropout and rf.weight_noise.

[albertz]

Still WIP. Well, the parametrization API is basically already ready. But I want to add some tests for it. Also I want to add weight dropout and variational noise already. Maybe also weight norm. Similar as here:
https://pytorch.org/docs/stable/_modules/torch/nn/utils/parametrizations.html
https://pytorch.org/docs/stable/generated/torch.nn.utils.parametrizations.weight_norm.html#torch.nn.utils.parametrizations.weight_norm

I'm probably not squashing all commits together but leave them as separate commits. This PR here is mostly just for easier reviewing, and maybe also discussions. (Although the main design questions are already being discussed in #1518.)

[albertz]

Ok, all done, at least weight_dropout and weight_noise. I think this is enough for this PR.

[albertz]

I already merged this because I want to use this in some experiments, and I think we can just directly fix any issues in the code in follow-up commits/PRs. I think we even can still do changes to the API for a short while, as long as no-one else is using this.

So, please anyway review/comment.