Fix training mode outputs for roberta and distilbert mlm models

[jimypbr]

What does this PR do?

• In training mode for MLM models we don't want to return the logits for performance.
• Distilbert and Roberta were only checking for if labels is None, but this meant that in evaluation the training path was being taken. Fix this by checking instead for if self.training.

[HuggingFaceDocBuilderDev]

The documentation is not available anymore as the PR was closed or merged.