Currently, empty Tensors may not have a Storage:

>>> x = torch.Tensor()
>>> x.storage() is None
True

I propose that we make every Tensor, even empty ones, have a Storage. I don't think this will require any changes to TH/THC (just pytorch).

This has some advantages:

• We often have code that assumes a valid storage. These bugs are hard to catch because they only occur with empty tensors.
• Every CUDA tensor will have an associated device. Currently, empty CUDA tensors may not be on any device