… on separate devices.

If either one of the metric device or the update input device
is a GPU, this commit will put the other one on GPU.

The comparison with self._device was not possible because it
can be created with `torch.device("cuda")` which is not equal
to `torch.device("cuda:0")` which is the device of a tensor
created with `torch.device("cuda")`. This change will have
a bigger performance hit when self._kernel is not on the same
device as y_pred as it will need to be moved onto y_pred's
device every time update() is called.

…l y_pred tensors are on the same device

…ensor with mypy

…elected device

Co-authored-by: vfdev <[email protected]>