Code Cleanup

[sdtblck]

Basically want the code to be easier for new devs to get their head around. Some suggested steps we could take toward this:

• refactor repo organisation (specifically which modules are in megatron/model/language_model.py / megatron/model/gpt2_model.py / megatron/model/transformer.py feels pretty arbitrary, I'd like the organisation here to be clearer so you can find the module you want more easily when making changes.)
• Get rid of unused codepaths and reduce duplication. (I think there are some modules that are never used still, for example, the Pooler class in megatron/model/language_model.py`
• Reduce complexity - if there's any way to de-spaghettify this code it should be done. E.G when instantiating a model without model parallel, we first call GPT2Model which then get_language_model which then calls TransformerLanguageModel which then calls ParallelTransformer... Dumb as hell.

[sdtblck]

I think with the latest pull from #269 we can call this done.

Lots of terrible megatron code has been removed, and we now have to maintain only a single model. Most unused codepaths are removed.