Remove __array__

[asmeurer]

This makes it raise an exception, since it isn't supported by the standard (if
we just leave it unimplemented, then np.asarray() returns an object dtype
array, which is not good).

There is one issue here from the test suite, which is that this breaks the
logic in asarray() for converting lists of array_api_strict 0-D arrays. I'm
not yet sure what to do about that.

Fixes #67.

This is built on #68, which can and should be merged separately (it isn't controversial).

[betatim]

In scikit-learn's tests we convert array API arrays to Numpy arrays in order to compare their values to some reference array. How would we do that if we remove __array__ here? Other libraries have .tonumpy() and the like, but I think array-api-strict doesn't? Should we add something like that as part of this removal?

[asmeurer]

Testing is a valid use-case. You could use dlpack if you wanted to stay within the standard. Although I'm still not clear if this is a good idea or not.

[betatim]

Good point. We already have custom code to convert from the various array libraries to Numpy. Using

x = xp.asarray([1,2,3], device=...)
np.from_dlpack(x)

would maybe simplify that code and remove the need for __array__.

[thomasjpfan]

Although I'm still not clear if this is a good idea or not.

I think this is a good idea for array-api-strict. Libraries that need to do any array conversion should only depend on __dlpack__.

(Although this PR is breaking tests in this repo)

[asmeurer]

To clarify a note from the OP: this has stopped working

>>> import array_api_strict as xp
>>> xp.asarray([[xp.asarray(1), xp.asarray(2)]])
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/Users/aaronmeurer/Documents/array-api-strict/array_api_strict/_creation_functions.py", line 97, in asarray
    res = np.array(obj, dtype=_np_dtype, copy=copy)
          ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/aaronmeurer/Documents/array-api-strict/array_api_strict/_array_object.py", line 138, in __array__
    raise ValueError("Conversion from an array_api_strict array to a NumPy ndarray is not supported")
ValueError: Conversion from an array_api_strict array to a NumPy ndarray is not supported

because it was implicitly relying on np.asarray() calling __array__ on nested array_api_strict arrays.

[asmeurer]

Thinking about dlpack, should we be doing anything with the device emulation in __dlpack__? https://data-apis.org/array-api/latest/API_specification/generated/array_api.array.__dlpack__.html#dlpack Maybe we should disable __dlpack__ when the device isn't the default CPU_DEVICE? @betatim

[betatim]

I thought dlpack now allows you to move arrays from one device to another? For example to support the common case of "i would like to convert this array from library foo to a numpy array". Is that right? If yes, it could be cool to support this for the "not CPU devices" here. If only because (in scikit-learn) we convert things back to Numpy arrays during tests.