Created by: suchenzang
Tackling https://github.com/facebookresearch/metaseq/issues/165
Only checkpoints to work with here are with OPT models so far.
More cleanup around unused args / logic in preparation for reducing complexity around configuration / checkpointing logic.
- Removed all normformer args & related classes (sync ln)
- Removed stable embeddings
- Removed FeedForwardNetwork class