Change transforms in RoBERTa into classes

[joecummings]

Currently, transforms in the RobertaBundle are defined as anonymous lambda functions. These are not pickleable and cannot be imported for use anywhere else.

Ex proposal:

lambda: T.Sequential(
        T.SentencePieceTokenizer(urljoin(_TEXT_BUCKET, "xlmr.sentencepiece.bpe.model")),
        T.VocabTransform(load_state_dict_from_url(urljoin(_TEXT_BUCKET, "xlmr.vocab.pt"))),
        T.Truncate(510),
        T.AddToken(token=0, begin=True),
        T.AddToken(token=2, begin=False),
    ),

-->

class RobertaTransform:
     def __init__(self, truncate_length=510):
         self.transform =  T.Sequential(
              T.SentencePieceTokenizer(urljoin(_TEXT_BUCKET, "xlmr.sentencepiece.bpe.model")),
              T.VocabTransform(load_state_dict_from_url(urljoin(_TEXT_BUCKET, "xlmr.vocab.pt"))),
              T.Truncate(truncate_length),
              T.AddToken(token=0, begin=True),
              T.AddToken(token=2, begin=False),
          ),

    def __call__(self, text):
        self.transform(text)

[Nayef211]

Hey I would recommend taking a look at #1482 where we actually moved away from using classes to using scriptable sequential container to create the transform. The main benefit from this was as follows

As we added necessary primitive transforms, along with support of #1481 for text transforms, we can easily avoid overhead of maintaining a separate class for model transform. Instead in this PR, we argue that this can be accomplished in much more sustainable way by directly constructing sequential transform as part of Model Bundle object.

As such do you think we want to migrate back to using classes for our model transforms?