Add a condition for checking labels

[hrxorxm]

What does this PR do?

This PR adds stability by checking whether it has labels before returning it in prediction_step of Seq2SeqTrainer
There is a similar condition in prediction_step of Trainer

if has_labels:
    labels = nested_detach(tuple(inputs.get(name) for name in self.label_names))
    if len(labels) == 1:
        labels = labels[0]
else:
    labels = None

Before submitting

• Did you read the contributor guideline,
  Pull Request section?

Who can review?

For trainer : @sgugger

[sgugger]

LGTM, thanks a lot for fixing!