[s2s]Use prepare_translation_batch for Marian finetuning (#6293)

Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

examples/seq2seq/utils.py

@@ -146,7 +146,9 @@ class Seq2SeqDataset(Dataset):
         return SortishSampler(self.src_lens, batch_size)
 
 
-class MBartDataset(Seq2SeqDataset):
+class TranslationDataset(Seq2SeqDataset):
+    """A dataset that calls prepare_translation_batch."""
+
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
         if self.max_source_length != self.max_target_length: