refactor superclass of MaskedLMModel (#2170)

Summary:
masked_lm is actually encoder-only model

# Before submitting

- [ ] Was this discussed/approved via a Github issue? (no need for typos, doc improvements)
- [x] Did you read the [contributor guideline](https://github.com/pytorch/fairseq/blob/master/CONTRIBUTING.md)?
- [x] Did you make sure to update the docs?
- [ ] Did you write any new necessary tests?

## What does this PR do?
Class Refactor

## Did you have fun?
Make sure you had fun coding �
Pull Request resolved: #2170

Reviewed By: ngoyal2707

Differential Revision: D21725071

Pulled By: myleott

fbshipit-source-id: 75fd36008f3e3425f8f5180472734394046dfb77

Author: xu-song

fairseq/models/masked_lm.py

@@ -11,7 +11,7 @@
 
 from fairseq import utils
 from fairseq.models import (
-    BaseFairseqModel,
+    FairseqEncoderModel,
     FairseqEncoder,
     register_model,
     register_model_architecture,
@@ -28,15 +28,14 @@
 
 
 @register_model('masked_lm')
-class MaskedLMModel(BaseFairseqModel):
+class MaskedLMModel(FairseqEncoderModel):
     """
     Class for training a Masked Language Model. It also supports an
     additional sentence level prediction if the sent-loss argument is set.
     """
     def __init__(self, args, encoder):
-        super().__init__()
+        super().__init__(encoder)
         self.args = args
-        self.encoder = encoder
 
         # if specified then apply bert initialization on the model. We need
         # to explictly call this to make sure that the output embeddings