Parallelformers

- Fix some multiprocessing bugs reported from https://github.com/tunib-ai/parallelformers/issues/32
- Add OPT models from https://github.com/tunib-ai/parallelformers/issues/30
- Add a guide for multiprocessing error in README.md

Fix gpu overallocation issue by adding deallocation before forward

Remove redundant operations

v.1.2.2
- [17] Fix performance issue about random sampling

[16] Remove assertion to force cpu usage

- [4] Add GPTJ
- [14] Add MegatronBert

Issues
- [4] Backward compatibility patch
- [5] Bug about `AlbertModel`

Patches
- [7] Backward compatibility patch
- support transformers from 4.2.0
- [6] Fix bug about `AlbertModel`