fix the crashing when there are zero-size inputs

[yongbinfeng]

PR to fix the crashing when some of the PyTorch inputs have zero size. One example here:

triton-inference-server/server#5972

The error seems tbe be because of Torch::from_blob function for copying inputs not working when the tensor is zero-size. The fix is to add the special handle in this case, directly creating a zero-size tensor, instead of calling the PyTorch function.

Tested and it seems to have fixed the crashing.

[Tabrizian]

Thanks for the contribution! I think I just need to add some testing and run it through the CI otherwise it is looking good.