Do we really need a `training_losses` fallback?

[ablaom]

We already have supports_training_losses. It also seems that documentation does not require the fallback, but this should be checked.