std::mdspan [C++23] conversion for pybind11::array

[karna48]

I have made a demo for explicit conversion of pybind11::array to std::mdspan.

  typename T, 
  typename Extents=std::extents<size_t, std::dynamic_extent>, 
  typename PyArray=py::array
>
std::mdspan<T, Extents> array_to_mdspan(PyArray &m) 

However I would love to implement implicit conversion so I could define pybind wrappers for functions like:

using extentsN = std::extents<size_t, std::dynamic_extent>;
void compute(std::mdspan<const float, extentsNx3> data_in, std::mdspan<float, extentsN> data_out);

PYBIND11_MODULE(pymdspan, m) {
    m.def("compute", compute);
};

Questions:

1. Maybe the conversions should be from pybind11::buffer not to require numpy.h ?
2. Which pybind11 conversion is a good start to look at, maybe std::vector or Eigen::Tensor?
3. If the mdspan should be returned from a function, which abstraction to prefer?

I'm really not sure where to start . But I think it should be easier than Eigen::Tensor because it is always a non-owning view.
So, please, if you could give me some pointers and suggestions.

Known issues:

• mdspan can handle strides, but I did not implement them yet nor I do check for them in py::array