Implement GPU/CPU parameter in QAPipeline()

[fmikaelian]

I think I understand. You are saying training the model on GPU and after that send it to CPU to make predictions.

It is already possible to do it (I did it with the CPU version trained on SQUAD), but it is not that transparent. After training the reader on GPU, you only need to run the code:

qa_pipeline.reader.model.to('cpu')

But we can also implement inside a method inside QAPipeline that does that, let's say to_cpu():

def to_cpu():
     self.reader.model.to('cpu')

Originally posted by @andrelmfarias in #133 (comment)