Switch to Pytorch's automatic mixed precision

[s-jse]

Pytorch 1.7 added support for mixed precision training and evaluation (https://pytorch.org/docs/stable/amp.html?highlight=amp#module-torch.cuda.amp).

This will be especially useful for speeding up inference for both semantic parsing and paraphrase generation. Also, training with mixed precision -if experiments show that it does not degrade accuracy- can help speed up our pipeline.