Compare PiPPy with MegatronLM performace

[pbelevich]

Possible scenarios:

• Compare plain PiPPy with plain Megatron
  • May be hard to implement
• Compare HF MegatronBERT with PiPPy-wrapped HF BERT
  • Choose a task (ForMaskedLM, ForCausalLM, ForNextSentencePrediction, ForPreTraining, ForSequenceClassification, ForMultipleChoice, ForTokenClassification, ForQuestionAnswering)
  • May depend on Mirror HF examples with PiPPy #301
  • Investigate if it's a fair comparison