Add domain modules trained End-to-End

[bdvllrs]

Allow for some domain modules to be trained end-to-end with the global workspace.

This brings some breaking changes:

1. DomainModule.compute_loss and DomainModule.compute_*_loss now require an 3rd
   parameter raw_target: Any that stores the raw domain input (before being encoded).
   This is usefull for unimodal losses that require the actual inputs to compute the loss.
2. GWLossesBase.step requires a new first argument raw_data: RawDomainGroupsT to
   pass the raw_targets to the domain modules.

1. needs to be changed in all projects that implement a DomainModule (every project).
2. has probably less impact as most project won't redefine a Loss module.

[RolandBERTINJOHANNET]

Once this is merged all we have to do on our personal projects is add that raw input parameter in the function calls and also the first arg for loss.step() ?

Shouldn't be too hard

[bdvllrs]

I don't think loss.step will be a problem, but yes for the domain modules loss function definitions.
I'm still missing to update the shimmer tutorial to reflect this change.

[bdvllrs]

Should be all good now.

[RolandBERTINJOHANNET]

yes, yes.

[bdvllrs]

After some internal discussion, we removed the End2EndDomainModule in favor of the freeze and unfreeze methods.

DomainModule stay frozen by default.

[RolandBERTINJOHANNET]

goood