_scipy_fft_backend has global side-effects

[peterbell10]

Here is a fairly minimal reproducer:

In [1]: import mkl, numpy as np
   ...: import mkl_fft
   ...: from mkl_fft._scipy_fft_backend import fft as scipy_fft
   ...:
   ...: x = np.random.rand(100, 100, 100).astype(np.cdouble)
   ...: mkl.set_num_threads(1)
   ...: %timeit scipy_fft(x, workers=8)
   ...: 
   ...: mkl.set_num_threads(8)
   ...: %timeit mkl_fft.fft(x)
320 µs ± 420 ns per loop (mean ± std. dev. of 7 runs, 1000 loops each)
3.36 ms ± 1.38 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)

I would expect the mkl_fft.fft call to use 8 threads and so be as fast as scipy_fft with workers=8. What actually happens is that scipy_fft set the fft domain thread count to 1 and the domain has higher precedence than the global thread setting.

mkl_fft/mkl_fft/_scipy_fft_backend.py

Lines 162 to 163 in 4d8cc2a

	n_threads = _hardware_counts.get_max_threads_count()
	mkl.domain_set_num_threads(n_threads, domain='fft')