argmax and argmin give incorrect results on option types when axis=None

[agoose77]

Version of Awkward Array

1.5.0

Description and code to reproduce

Consider

>>> x = ak.Array([1, 2, 3, None, 4])
>>> ak.argmax(x)
3

What is happening is that ak._util.completely_flatten is dropping the None values, consequently shifting the position of the elements before calling np.argmax.

[jpivarski]

Right. This could be fixed by having argmin/argmax with axis=None first apply ak.fill_none with axis=-1 and a fill value that is -np.inf for argmax and np.inf for argmin, before running it through ak._util.completely_flatten.

>>> ak.argmax(ak.fill_none(x, -np.inf, axis=-1))
4

This converts the numerical data into floating point numbers, which doesn't break argmin/argmax (positions of minimum/maximum are unchanged), but it's an unnecessary computation. A more sophisticated version would use np.iinfo(dtype).min and .max instead of -np.inf and np.inf, but an array could contain multiple dtypes. It's maybe best for a first fix to just use -np.inf and np.inf.

[ianna]

or by using axis=-1:

>>> x = ak.Array([1, 2, 3, None, 4])
>>> ak.argmax(x, axis=-1)
4

[jpivarski]

I had forgotten about this, but it's an important bug to fix promptly. I applied the easy fix described above in both v1 and v2.