port scalarmath tests

[ev-br]

Make minimal modifications to dtype/scalar type machinery to get scalar math tests run. Lots of xfails/test skips for now, but I think this is more or less it for dtypes. This PR is against the gh-12 branch, these two PRs are a single "stack", as it were.

Next up is ufuncs, esp those which are also ndarray methods.

[lezcano]

A general question. Can you remind me why we need to explicitly have a base member within _ndarray? What use cases does it cover?

[lezcano]

Why not simply return asarray(zeros)?

[ev-br]

At some point we'll need to rationalize these two forms, agree.
Basically, asarray is anything array-like in, array out; here we explicitly construct the tensor, so my fingers naturally typed this line. The line above, with asarray(self._tensor.imag), should be changed to follow line 94.

[lezcano]

I mean, asarray(Tensor) has the same semantics as _from_tensor_and_base(Tensor, None), so we can decide to always prefer the first one over the latter one for conciseness and consistency.

[ev-br]

Not always

In [2]: a = np.zeros(3)

In [3]: np.asarray(a) is a
Out[3]: True

[lezcano]

note that I suggested doing so when using torch.Tensors.

[ev-br]

A general question. Can you remind me why we need to explicitly have a base member within _ndarray? What use cases does it cover?

In numpy it's, AFAIU, a python side of the pointer chain for ownership of the array data:

In [1]: import numpy as np

In [2]: a = np.arange(4)

In [3]: a[::2].base is a
Out[3]: True

In [4]: a[::2][1] = 8

In [5]: a
Out[5]: array([0, 1, 8, 3])

So if arr is a view of some other array, a.base points to it. (Here view is numpy view, not pytorch'es).
Since it's public in numpy, why would we not keep it as well.

[ev-br]

Review comments addressed in 2b95ac2

[lezcano]

About #16 (comment), sure, but this tracking also happens in torch.tensor()._base, so why not just return that one?

[lezcano]

We should discuss asarray vs _from_tensor_and_base (probably remove the latter one) and potentially removing the base attribute from the wrapper. Now, we don't need to block this PR on that.

[ev-br]

Thanks for the review Mario!
Merged in f63c6bb