Use native sizes by default when parsing buffer format string

[JobJob]

Ahh oops. So, apparently you're supposed to use native sizes by default when parsing the buffer protocol format strings. I did the opposite in #487.

Current master

using PyCall
py"""
import numpy as np
"""
nparr = py"np.array([1,1,1,1])"o
nparr["dtype"]  # PyObject dtype('int64')
convert(PyAny, nparr)

returns Int32[1,0,1,0] *
It should just be Int[1,1,1,1]

This actually caused me a bug in some relatively large distributed code, that I found very hard to track down. So karmically, pretty interesting.

At least it didn't make it to a release.

* The [1, 0, 1, 0] is also in part due to a bug in f_contiguous for 1d arrays when stride != sizeof(T) - I will post a PR with a fix for that next week, I need to make some other changes to get the tests right for that. This is the minor positive to come out of this.

[stevengj]

What is the AppVeyor failure?

[JobJob]

Just an fyi: with the appveyor failures before, it seems like the conda x64 environments might not have actually been 64-bit builds of Python. The tests were failing because the dtype of a np.array([1,2,...,10]) in python was int32, not int64, but in julia Int was Int64 on those environments.