MAINT: Use PySlice_GetIndicesEx instead of custom reimplementation

[njsmith]

This has the side effects of:

• changing several IndexError exceptions into TypeErrors
• allowing slices like arr[False:True] as equivalent to
  arr[0:1] (because now we're using Python's logic for interpreting
  slices, and Python is happy with treating bools as integers in integer
  contexts).

It also deletes almost 100 lines of code :-).

[njsmith]

CC @seberg

[seberg]

Sounds good to me, allowing booleans in slice context seems not so bad. I think I would wait with putting this in until the dust settled on 1.11 though.

[charris]

Two of the errors (warnings) look legitimate.

[njsmith]

Ugh, the warnings are totally not legitimate... on py2, the first argument is typed as PySliceObject*, and on py3 it's typed as PyObject*, so the implicit cast is fine and there's no way to make both versions happy at the same time. But we need to do something to make the tests work; I will ponder.

[juliantaylor]

nothing a little macro magic can't solve, see e.g. nditer_pywrap.c for the same issue

[juliantaylor]

that solution could be put into the 3kcompat header if we need it more than once

[njsmith]

@juliantaylor: oh hey, and those nditer_pywrap.c usages are wrong anyway, because they call the terrible PySlice_GetIndices instead of the functional PySlice_GetIndicesEx. (in particular the calls in nditer_pywrap.c assumes that it sets an exception on error, which is not true...)

[njsmith]

Fixed / rebased / squashed, so should be good to go (hopefully travis agrees).

I can see @seberg's argument that we might want to wait on merging this until #7162 is resolved.

[charris]

Thanks Nathaniel.