API: remove ndarray.base

[ev-br]

closes gh-47

Base tensor relationship is tracked via pytorch'es tensor._base already, so remove ndarray.base attribute.
Checking an wrapper array base becomes then

>>> a.get()._base is b.get()

instead of numpy's

>>> a.base is b

This is the second PR in the "stack" based off gh-70.

EDIT : a.get() is gone, so it's a.tensor._base now.

closes gh-47

[lezcano]

In this clean-up, we can also remove the .get() method (now it's just ._tensor (or .tensor if you want, no need to make that member private, it's a bit unpythonic IMO) and ndarray._from_tensor(x) can be merged into the __init__ so that we can write the more natural ndarray(t).

[ev-br]

FWIW, I'd rather keep only having an argument-less __init__ or maybe even no __init__ at all, to avoid mixing up with https://numpy.org/doc/stable/reference/generated/numpy.ndarray.html which is very much not recommended for a long time now but is still used in old scripts here and there.

W.r.t. public or private tensor attribute, I suppose my instincts are from a library maintainer perspective : anything which is not explicitly private will get (ab)used in the wild and will cause backcompat issues down the line. Are we not worried about future backcompatibility?

[lezcano]

Note that this ndarray is not a public API but an intermediary artefact used by a compiler, so we should write it in whichever way we find easier.
Same for __init__ really. Note that it makes sense for np.ndarray(...) not to work because you can construct an array in a thousand ways. Our arrays are only thin wrappers around tensors, so we don't have that issue.
If you want to be extra sure that it's being used as expected, you can put an assert t is None or ìsinstance(t, torch.Tensor) in the __init__ constructor.

[ev-br]

OK, thanks. Having no public API does make it easier indeed.

FTR, I still believe constructing a thing from a tensor is a prime example for a from_tensor alternative constructor. That said, this is not my hill, not today anyway, and the last commit ditches the get() getter, adds ndarray.__init__(self, t: torch.Tensor), and makes ndarray.tensor public.

[lezcano]

Nice clean-up!