Add some fixes

[Foxglove144]

1. Add requirements.txt.
2. Move tokenized input to CUDA in translation.py and token_classification.py.
3. Fix bug in printing predictions from decoded output in token_classification.py.
4. Change number of samples per batch during training and eval from 16 to 8 to prevent CUDA out of memory error in summarization.py.
5. In masked_language_modeling.py, use the evaluate() method to print model perplexity.

[muellerzr]

Thanks! Seems like good improvements to me, but left some comments

[Foxglove144]

Moved tokenized input to correct device automatically in translation.py, token_classification.py, sequence_classification.py and translation.py

[muellerzr]

I'm going to leave this PR open, however the goal of this isn't to be device-agnostic, but to be very simplified examples. Which also includes not having accelerate involved in the script itself.

We can keep this up for future people to reference however