Don't let pytorch __call__ implementation for modules circumvent our typing

[glencoe]

It would be great to have typing annotations on our forward functions be mirrored in the type annotations for call. E.g, LSTM actually returns tuple[Tensor, tuple[Tensor, Tensor]] and our forward function says so, but that information is overriden by pytorch's implementation of nn.Module.__call__. So either we should override that in our own modules and add the type annotation or we should find a way to programmatically propagate the type annotation from forward to __call__

[julianhoever]

@glencoe I will have a look at it. Unfortunately, I have not yet started to integrate our own tensors into our modules. That's why I haven't encountered the problem yet. But that is the next step. Then I will try to get a fix for it.

[glencoe]

@julianhoever the issue arises from how pytorch treats their type annotations. It's not a bug ;) and not related to introducing our own tensor types. It's not urgent, i just added so we do not forget it.