[docs] Tensor.new is not documented

[shekhovt]

Docs master has in http://pytorch.org/docs/master/tensors.html:

new(*args, **kwargs)
Constructs a new tensor of the same data type.

Missing documentation of arguments. From some other pages it can be figured out that new is preserving the storage and even the device number and can take a size argument. Would be good to have complete specs.

[soumith]

fixed via #3954

[ssnl]

tensor.new became not documented in 0.4

[zou3519]

It's a legacy method so I chose not to document it. If we want it we can put it back (and add a deprecation warning)

[ssnl]

@zou3519 We should still put it there. It is used in a lot of community repo (and our code base too). People will want to look up the doc.

[PiotrCzapla]

I ran into this issue analyzing some community code. It would be good to know how to replace it with a method that isn't deprecated.

[soumith]

@PiotrCzapla torch.empty would be a replacement for .new https://pytorch.org/docs/master/torch.html#torch.empty

[vadimkantorov]

One way why new is still used is that it behaves a bit similar to _like - preserves device and dtype.

If new is still used, i propose it can construct scalar tensors, e.g. torch.rand(1).new(0.5), currently it doesn't work