Adds basic support for MPI + LAMBADA dataset

[ArulselvanMadhavan]

• I wanted to evaluate the FORTRAN implementation on the LAMBADA dataset. It would be nice to match the evaluation metrics reported here - https://huggingface.co/gpt2
• I have seen implementations from other language being published to HuggingFace. Do you have plans of publishing it there? I don't know much about it either. But, I am curious.
• Ended up adding basic MPI support.
• I wanted to implement the perplexity metric but the PR was getting bigger. So, I figured I'd share what I have. To implement perplexity, we need to calculate loss. To calculate loss, we have to truncate the sequence length to total - num_tokens_to_generate and use that as the target. Loss = NLL(generated_output, target); Then, Perplexity = exp(sum(losses)/num_generated_tokens)

Thanks for creating this project!

[certik]

Thanks for your interest! Yes, see #31 for uploading to HuggingFace or elsewhere online. I recommend you split this PR into smaller ones and in each implement just one feature, such as improvements to cmake, etc.