npy_format_descriptor may not be accurate for platform-dependent lengths of short, long, and long long

[EricCousineau-TRI]

Ran into this via an esoteric route: When tinkering with user-defined dtypes, it seems that implementation of np.ones uses np.copyto(dst, 1, casting='unsafe'), where the 1 gets interpreted as np.array(1, dtype=np.int64) (at least, on my platform).

When I tried registered a caster using npy_descriptor_format<int64_t>::dtype() using the type_num field, I realized its value was for NPY_LONGLONG, whereas NPY_INT64 corresponds to NPY_LONG.

Will try to see if I can create a failing test, and then submit a tentative fix PR based on that.

[EricCousineau-TRI]

Yup, could reproduce the issue, using the latest master (6d0b470):
EricCousineau-TRI/repro@de9c086#diff-9885f7c92e2882fb9307118ce30b0aa7

Output on my machine:

9   <-- What pybind11 reports as the dtype for `int64`
9
7
7
7   <-- What NumPy says is the dtype for the `NPY_INT64` alias

NumPy info (using it from Ubuntu 16.04):

>>> import numpy as np
>>> np.version.full_version
'1.11.0'
>>> np.version.git_revision
'4092a9e160cc247a4a45724579a0c829733688ca'

TBH, it'd have been nicer if NumPy had anchored the typedefs according to concrete sizes, and not platform-dependent types :/ Oh well.