mkl_fft.fft and mkl_fft.fftn leaks memory on real inputs

[oleksandr-pavlyk]

Reported via Intel Developer's Forum post.

Reproducer:

import numpy
import math
import gc


import numpy.fft as fft

while True :
	test=numpy.fft.fft2(numpy.ones((200,200)));
	del test
	gc.collect()

process memory increases unboundedly over time.