Better loss system for broadcast_loss

[bdvllrs]

if compute_loss returns None, it's skipped + Broadcast loss uses tr, dcy, cy and fused losses

[RolandBERTINJOHANNET]

In which situation would compute_loss return none ? Wouldn't we want an error in this case ?
Also I just want to check, the functions compute_[...]_loss are added in this PR so we can choose to have different computations based on the loss type ?

[bdvllrs]

In which situation would compute_loss return none ? Wouldn't we want an error in this case ?

The rationale is that a domain module can choose to never have one of the loss (e.g. to not participate to translations).
Choosing a loss coef to 0 is a user choice, setting one loss to None is a domain choice. It obviously has priority over a non 0 loss coef.

Also I just want to check, the functions compute_[...]_loss are added in this PR so we can choose to have different computations based on the loss type ?

Those were not added in this PR and have been part of shimmer for some times already, but it was not used by the broadcast loss. But basically yes, it's useful in case translation and demi cycles not use the same exact loss function.

[RolandBERTINJOHANNET]

OK makes sense :)

[bdvllrs]

I added documented the change and explicited what I just wrote. I agree, it wasn't clear originally.