Enable dndarray-torch.tensor operations

[ClaudiaComito]

the way we do it right now is e.g., ht.zeros(5).larray + torch.randn(5) and then wrap it into a (distributed) dndarray again. It would be less frustrating if this step were taken care of internally.

Issue created from a Mattermost message by @ClaudiaComito.

[dizcza]

I see two scenarios:

1. ht.zeros(5) + torch.randn(5) -> ht.DNDarray
2. ht.zeros(5).tensor() + torch.randn(5) -> torch.Tensor

The .(to|as)tensor function (or .view(torch.Tensor)) should cast a DNDarray to a torch.Tensor without copying the data, if possible (read if allocated on the same node), or optionally provide copy=True flag. Having such views onto PyTorch tensors means that inplace operations such as

a = ht.zeros(5)
a.tensor().add_(torch.arange(5))

modify the data in a.

[ClaudiaComito]

Thanks @dizcza ,

note that option 2. is possible already as:

ht.zeros(5).larray + torch.randn(5) -> torch.Tensor

(in the current main branch. The syntax in 0.5.0 is ht.zeros(5)._DNDarray__array + torch.randn(5) -> torch.Tensor its non-intuitiveness reflects the fact that we only used it internally so far)

You're totally right though that ht.zeros(5).tensor() is a more intuitive solution.

[mrfh92]

In principle very useful, but maybe not of highest priority (but only since there is some workaround).

(Reviewed within #1109)