C++ refactoring: _getitem_array implementation

[ioanaif]

• Content.__getitem__ in the right order
• EmptyArray._getitem_array
• NumpyArray._getitem_array
• RegularArray._getitem_array
• ListArray._getitem_array
• ListOffsetArray._getitem_array
• RecordArray._getitem_array handing the allow_lazy case
• IndexedArray._getitem_array
• IndexedOptionArray._getitem_array
• ByteMaskedArray._getitem_array
• BitMaskedArray._getitem_array
• UnmaskedArray._getitem_array
• UnionArray._getitem_array

[jpivarski]

Copying from Slack, Content.__getitem__ should be roughly:

1. check int
2. check slice
3. check string
4. check non-tuple iterable of strings
5. ...
6. check ak.Array, pull out its layout and call __getitem__ on that so that it can be handled by rule 7 or 8
7. check NumpyArray (to enter into _getitem_array)
8. check generic Content (other cases, like jagged arrays and missing values)
9. check tuple (for the generic _getitem_next)
10. check remaining Iterable by calling nplike.asarray on it, wrapping it in NumpyArray and passing it into __getitem__ again so that it gets picked up by rule 7 and goes into _getitem_array
11. TypeError

(roughly)

[jpivarski]

For testing, you can alternate between old-style and new-style arrays. The ak._v2.tmp_for_testing module defines v1v2_equal, which you can use to see if an old-style and a new-style array are exactly equal, and v1_to_v2/v2_to_v1 for conversions.

When your project is done, we'll be removing the ak._v2.tmp_for_testing module as well as the old-style arrays, so while they're useful for ensuring that your new functions are compatible with the old ones, the final tests will have to be written in such a way that we can remove the old stuff and not lose the tests.

You can do this by developing with direct comparisons (v1v2_equal), then make the final tests look like

result = your_v2_array.something()
assert ak.to_list(result) = [...]
assert result.form.tolist() == {...}
assert v1v2_equal(result, v2_to_v1(your_v2_array).something())

where you get [...] and {...} from the old-style array, and they're plain Python data. When we remove the old-style arrays, we can remove the last assertion, but the other two completely check the correctness of the array (values and detailed type).

[jpivarski]

All Content subclass constructors (e.g. calling ListArray, IndexedArray, etc.) need to pass in the appropriate identifiers and parameters. We can't define identifiers yet, but I'll make utility functions to do the right thing when they're not all None (and pass through None for now). Passing through or creating empty parameters is very important, so the next PR should get started doing that.

Additionally, I'll make the very early (low numbered) tests start testing the new code soon.

This PR was in very good shape before I started messing with it. A few things were simplified by using kernels, rather than NumPy functions, which is something that we can do now, due to #966.

I'll try to get some underlying infrastructure done before you come back to this next week.