PYBIND11_NUMPY_DTYPE does not work on structure containing array

[bmerry]

Issue description

Using PYBIND11_NUMPY_DTYPE on a structure member which is an array (either a language array or std::array) leads to a compilation error:

pybind11/numpy.h:1027:9: error: ‘format’ is not a member of ‘pybind11::format_descriptor<int [2]>’
         ::pybind11::format_descriptor<decltype(std::declval<T>().Field)>::format(),           \
         ^

Reproducible example code

#include <pybind11/pybind11.h>
#include <pybind11/numpy.h>
#include <array>
#include <cstdint>

namespace py = pybind11;

struct foo
{
    std::int32_t array[2];
};

PYBIND11_PLUGIN(arraytest)
{
    PYBIND11_NUMPY_DTYPE(foo, array);

    py::module m("arraytest", "doc");
    return m.ptr();
}

Compiled with GCC 5.4 with g++ -std=c++14 -c pybind11-array.cpp plus -I options for Python and pybind11, and pybind11 version 2.1.1

[jagerman]

It's an intentional, explicit omission: not because it's impossible, but just because no one has done it yet.

[aldanor]

Well, almost... we wouldn't mark arrays as is_pod_struct, but rather specialise npy_format_descriptor and format_descriptor. In fact, we do specialise those for char[N] and std::array<char, N>, mapping them to S dtype which is admittedly slightly hacky (since it overlaps with int8_t[N], if we were to map arbitrary fixed-size arrays).

[bmerry]

I don't think it's likely to conflict with int8_t[N], because int8_t will be signed char and char is a distinct type from both signed char and unsigned char.

I might have a go at it if I get time, although it won't be this week.