ARROW-4181: [Python] Fixes for Numpy struct array conversion

[pitrou]

This fixes two issues:

• object fields inside numpy structs should be allowed
• buggy ndarray indexing if stride % itemsize != 0

[kszucs]

@pitrou You can test edge cases with hypothesis, like the following:

import hypothesis as h
import hypothesis.extra.numpy as npst

supported_scalar_dtypes = (npst.boolean_dtypes() | 
                           npst.integer_dtypes() | 
                           npst.unsigned_integer_dtypes() | 
                           npst.floating_dtypes() | 
                           npst.datetime64_dtypes())

@h.given(
    npst.arrays(npst.nested_dtypes(supported_scalar_dtypes), shape=(5,))
)
def test_nested_struct_array_from_numpy(numpy_array):
    pa.array(numpy_array)

Or to test converting to pylist:

import hypothesis as h
import pyarrow.tests.strategies as past

@h.given(
    past.arrays(past.nested_struct_types())
)
def test_nested_pyarrow_array_to_pylist(pyarrow_array):
    pyarrow_array.to_pylist()

And run with:

pytest -sv --enable-hypothesis --hypothesis-show-statistics --hypothesis-profile=dev

or to print the generated examples:

pytest -sv --enable-hypothesis --hypothesis-show-statistics --hypothesis-profile=debug

[pitrou]

When fixing a specific bug, I think I'd rather rely on manual testing and know exactly what's being tested.

[kszucs]

Agree, just wanted to inform You about a handy way to discover bugs :)

[xhochy]

+1, LGTM