Writing a cupy array caster for pybind11 bindings

[SimeonEhrig]

During the development of PR #15 we recognized, that we cannot use the numpy array support of pybind11 for cupy arrays. The reason is, that both arrays uses different apis.

• cupy : https://numba.pydata.org/numba-doc/latest/cuda/cuda_array_interface.html#cuda-array-interface
• numpy: https://docs.scipy.org/doc/numpy-1.13.0/reference/arrays.interface.html#__array_interface__

Therefore we need to implement a custom pybind11 caster to support a nice interface for the C++ and Python side.

Please implement the custom caster. The C++ interfaces should look like this:

void compute(mycupy::arrray_t<float> x, mycupy::arrray_t<float> y){
  kernel<<<1,1>>>(x.ptr, x.size, y.ptr, y.size);
}

This documentation contains useful information about, how a Python class is organized internal: https://docs.python.org/3/reference/datamodel.html

[SimeonEhrig]

Update: During the development of the cupy caster, we got new experience using pybind11, which changes some of the design decisions.

• Python to C++: By default, pybind11 passes arguments by copy from the Python to the C++ side. With a custom caster, it is also possible to pass a reference via pointer. But the breaks the expected behavior of the pybind11 api. Therefor we introduce the Cupy_Refclass, which is a reference to a cupy array and pass this wrapper class by reference to the C++ side.
• C++ to Python: For the other way, we can use directly a cupy array, because the ownership of cupy array needs to be moved to the Python side after leaving the C++ function.

[SimeonEhrig]

Done in #22