Porting OPT-175B to this framework

[CoderPat]

Currently, this framework only supports BLOOM models with DeepSpeed
However, various results have shown that OPT-175B is better than BLOOM (citation needed).

Ideally, we would like this framework to support OPT-175B as well.
The main work would be in the form of loading OPT-175B in an huggingface model and sharding with deep-speed. This would probably require a machine with loads of RAM (but hopefully without the need for 8 GPUs)
I've started an issue in the original repo for this huggingface/transformers-bloom-inference#39