Fix KeyError when initialize the model with ignore_mismatched_sizes=True

[tricktreat]

What does this PR do?

KeyError is thrown when the model is initialized from pre-trained weights with ignore_mismatched_sizes=True. Reproduced as follows:

>>> import transformers
>>> transformers.BertModel.from_pretrained("bert-base-cased", ignore_mismatched_sizes=True)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/shenyl/miniconda/envs/cuda111/lib/python3.8/site-packages/transformers/modeling_utils.py", line 1882, in from_pretrained
    model, missing_keys, unexpected_keys, mismatched_keys, error_msgs = cls._load_pretrained_model(
  File "/home/shenyl/miniconda/envs/cuda111/lib/python3.8/site-packages/transformers/modeling_utils.py", line 2003, in _load_pretrained_model
    and state_dict[checkpoint_key].shape != model_state_dict[model_key].shape
KeyError: 'bert.embeddings.LayerNorm.weight'

The cause is that the key modified by function _fix_key is not found in state_dict.

The solution is to use the original loaded keys when finding the mismatched keys.

[HuggingFaceDocBuilderDev]

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

[sgugger]

Confirming this fixes the issue, thanks a lot for your PR!