training_forward assumes input has .copy() method for single GPU

[MikeScarp]

Common bugs:

1. Tensorboard not showing in Jupyter-notebook see issue 79.
2. PyTorch 1.1.0 vs 1.2.0 support see FAQ

Describe the bug
After #532 , training with a DataLoader where batch does not have a .copy() method (such as tuple) will cause an exception. It doesn't seem to be the assumption that batch should always be a tensor, because we are passing it to transfer_batch_to_gpu, which does a lot of checking to handle different types differently.
Exception happens at https://github.com/williamFalcon/pytorch-lightning/blob/f2191b0cdf4305ae3a5ad2b1e404f99764a1a7c6/pytorch_lightning/trainer/train_loop_mixin.py#L293

To Reproduce
Steps to reproduce the behavior:

1. Install bleeding edge build
2. Train with an DataLoader where a batch is of type tuple with one GPU

Expected behavior
I would expect the copy to only happen when batch is of type tensor, or possibly apply the copy to elements inside the tuple if it is a tuple.

[williamFalcon]

@MikeScarp good catch, silly oversight on our part. Mind submitting a PR?

[giovannipcarvalho]

It might not be a tuple either. I'm experiencing the same with torchtext's Batch class. In my case, data is already on the GPU.

[awaelchli]

Looks to me like the problem described by @MikeScarp was solved. Master branch does not contain the copy anymore.

What @giovannipcarvalho mentions is also discussed here: #1245
So this may be closed? @MikeScarp