Enable zero-size tensors (#3640)

Summary:
Pull Request resolved: https://github.com/pytorch/executorch/pull/3640

As title.

The approach is slightly different than in PyTorch Vulkan. Instead of binding no memory, we make a small allocation. The reason for this change is to account for the possibility that some zero size tensors are used as input but the output is not zero size. In that case we still need to be able to bind the zero size tensor to a shader.

Reviewed By: yipjustin

Differential Revision: D57450473

fbshipit-source-id: 753e0dedf40f4a6ee6153980159fce34b8c41b5e