bugfix: PytorchTrainer memory leak

[DNA386]

Detach the tensors returned in the training loops to ensure the computation graph can be cleared from memory properly after each batch.

[y-richie-y]

As far as I know, .item() already converts the loss to a python scalar, hence detaching the value from the computation graph, so .detach() is redundant. Once loss goes out of scope, it can be garbage collected. Can you provide more information or give an explanation as to why this resolves the memory leak?

[DNA386]

detatch not required with item()

[DNA386]

You're right, its not needed on the loss.
We do still need to detach y though (which was the actual source of the leak)