[BUG] `auto_move_data` does not work with `DataParallel`

[YannDubs]

🐛 Bug

In case your forward function is wrapped with auto_move_data it will not work with DataParallel because it will try to send the data to self.device which in dataparallel is always the main device.

i.e. the following won't work with accelerator="dp" (and probably also with "ddp"):

class Module(pl.LightningModule):
     ...
     
     @auto_move_data
     def forward(x):
         ...

     def training_step(self, batch, batch_idx):
         x = self.forward(batch[0])
         ...

The error comes from this line: https://github.com/PyTorchLightning/pytorch-lightning/blob/b190403e282cbcb71147c7b618654476b08578a5/pytorch_lightning/core/hooks.py#L646

self.device should probably be replaced by torch.distributed.get_rank() when torch.distributed.is_available() and torch.distributed.is_initialized()

[awaelchli]

Hi, thanks for reporting.
For this we need #6414, then it can work. I will update the PR soon to include the test for auto_move_data.