non-native endianness ndarrays cause a crash, instead of silent conversion to native endianness

[brandondube]

This is essentially the same issue as cupy/cupy#4288;

A basic example:

from astropy.io import fits
from mkl_fft import _scipy_fft_backend as mklfft

a = fits.getdata(any_fits_file)
mklfft.fft2(a)

Produces the exception:

~/miniconda3/envs/prysm/lib/python3.8/site-packages/mkl_fft/_scipy_fft_backend.py in fft2(a, s, axes, norm, overwrite_x, workers)
    187     x = _float_utils.__upcast_float16_array(a)
    188     with Workers(workers):
--> 189         output = _pydfti.fftn(x, shape=s, axes=axes, overwrite_x=overwrite_x)
    190     if _unitary(norm):
    191         factor = 1

mkl_fft/_pydfti.pyx in mkl_fft._pydfti.fftn()

mkl_fft/_pydfti.pyx in mkl_fft._pydfti._fftnd_impl()

mkl_fft/_pydfti.pyx in mkl_fft._pydfti.iter_complementary()

<__array_function__ internals> in copyto(*args, **kwargs)

TypeError: Cannot cast array data from dtype('complex128') to dtype('>f8') according to the rule 'same_kind'

mkl_fft version 1.3.0