Added automatic mixed precision (AMP) support for reader training from Haystack side

[antoniolanza1996]

Resolves #462.

[antoniolanza1996]

I have added all the stuff to enable AMP from Haystack side.
However, there are still problems from FARM side, as you can see from this error message:

RuntimeError: Found param language_model.model.embeddings.word_embeddings.weight with type torch.FloatTensor, expected torch.cuda.FloatTensor.
When using amp.initialize, you need to provide a model with parameters
located on a CUDA device before passing it no matter what optimization level
you chose. Use model.to('cuda') to use the default device.

@tholor , can you fix it?

[tholor]

@antoniolanza1996 I finally had a chance to test this. However, I could not replicate the above issue. For me it seems to run correctly when executing:

reader = FARMReader(model_name_or_path="distilbert-base-uncased-distilled-squad", use_gpu=True)
reader.train(data_dir= "data/squad20", train_filename="dev-v2.0.json", use_gpu=True, use_amp="O1", n_epochs=1, save_dir="my_model")

Did you maybe load the FARMReader without "use_gpu=True"?

[antoniolanza1996]

@tholor you are right. I have re-run and now it works well.
Probably last time there was something that I didn't appropriately set, sorry for that.
Hence, I think you can merge this PR, isn't it?

[tholor]

Yep, we can merge then