Post 0.4 tasks to do

Higher Priority

Issue	PR
[feature request] match numpy behavior on zero-length and zero-dimension tensors
torch.matmul() behavior doesn't match docs	PR#6961
torch.arange type	PR#7016
[PyTorch] Add torch.astensor and deprecate torch.from_numpy	PR#7109

Lower Priority

Issue	PR
torch.tensor / Tensor.new_tensor should be able to take data + sparse dtype
[Feature request] A few things that numpy does that pytorch could have (means with floating types)