Fix unpickling issues with CUDA sub-attributes

[ottonemo]

Solves issues with, for example,

• criterion__weight = <cuda tensor>
• optimizer__state
• module__some_weird_parameter = <cuda_tensor>

Before, criterion was never considered a CUDA dependent
attribute which is also fixed.

[ottonemo]

I think there was also a bug where net.cuda_dependent_attributes_ was overwritten with the torch byte array of said attributes when unpickling. This is also fixed by introducing the intermediate state variable __cuda_dependent_attributes__.

[BenjaminBossan]

LGTM. Could you quickly comment on why that part is no longer required:

https://github.com/dnouri/skorch/blob/master/skorch/net.py#L1422

(or was it ever?)

[ottonemo]

I don't think it was ever needed. Stumbled upon it a while ago and used this occasion to get rid of it.

There's one problem though, what do we do about old pickled files? These will fail to load with this change since they are missing the __cuda_dependent_attributes__ state variable. We could load the old way and issue a warning that this fail will fail to load with the next release?

[ottonemo]

I've added a compatibility check so that we can direct users to this version of skorch if they have problems loading their old pickle files. In future versions this fallback can be removed.

[BenjaminBossan]

This looks good, just one minor thing: Could you mention in the CHANGES.md that the old pickle format is deprecated and users should re-pickle?

[ottonemo]

@BenjaminBossan done, please review again!