Add __array_ufunc__ and __array__function__ to ArrayIndex

[telamonian]

Fixes #95

Adds appropriate definitions for the __array_ufunc__ and __array__function__ to the ArrayIndex class. This makes it so that if a is a ndarray and b is a ArrayIndex, both a == b and b == a get dispatched to b.__eq__(a)

[asmeurer]

You should update the == tests in test_ndindex.py.

What are the other consequences of this? I don't want array index types to be treatable as arrays (see https://quansight.github.io/ndindex/type-confusion.html#integerarray-and-booleanarray). Array operations should in general not work on them.

[asmeurer]

Oh I didn't see your comment on the issue. Given that, this does seem like the best option.

[asmeurer]

Also I just merged #93 which updates that test, so you'll need to update this on top of that.

[asmeurer]

Add a comment here explaining why these are here.

[asmeurer]

This looks good. I would just like a comment explaining what this does. I guess it's more possible to override NumPy behavior than I thought. Now if only NumPy would allow a[idx] to be overridden so that we don't have to use a[idx.raw] for ndindex types.

[telamonian]

There's probably a NEP for that. Turns out there's a NEP for everything :)

[asmeurer]

I wonder if we should also at least define

def __array__(self):
    return self.array

Right now we have

>>> a = np.array([True, False])
>>> idx = ndindex(a)
>>> np.array(idx)
array(BooleanArray([True, False]), dtype=object)

We could also make it give an error.

    def __array__(self):
        raise TypeError(f"Cannot convert {self.__class__.__name__} to an array. Use .array instead.")

I'm not sure which way is better.

[telamonian]

ArrayIndex instances are not arrays. So they shouldn't just transparently cast to arrays.

They're hardly rules to live by, but I've found the zen of python to be good guidelines for building the simplest, most intuitable version of complex behaviors. Here I think the relevant parts are

"Explicit is better than implicit."

and

"There should be one-- and preferably only one --obvious way to do it."

So I think your second suggestion about raising a TypeError is good. But I'd go even further and say that the .array prop is superfluous and potentially confusing, and should be removed in favor of just .raw

[telamonian]

This looks good. I would just like a comment explaining what this does. I guess it's more possible to override NumPy behavior than I thought. Now if only NumPy would allow a[idx] to be overridden so that we don't have to use a[idx.raw] for ndindex types.

Honestly this does seem like a good idea. Maybe we really should consider opening a NEP about ndarray.__getitem__ dispatch. I opened an issue for this at #97

[asmeurer]

Sorry for not looking into your PRs. The CI was broken on ndindex because of an upstream numpy issue, but that's been resolved now that the 1.20 prerelease is out. I've pushed a commit here addressing my comment, and I will merge once the tests pass.