Release GIL in gradient search resampling

[djhoese]

Problem description

It was noted in recent EWA profiling that the Python Global Interpreter Lock (GIL) was not being released during processing which caused CPU usage to hover around 100% even though multiple dask workers were being used. Adding with nogil: to the proper spots in the Cython code improved execution time and increased CPU usage to ~700% usage for 8 workers. This same lack of releasing the GIL was noticed in the gradient search code. Specifically here:

pyresample/pyresample/gradient/_gradient_search.pyx

Lines 124 to 133 in c930465

	one_step_gradient_search_no_gil(data,
	src_x, src_y,
	xl, xp, yl, yp,
	dst_x, dst_y,
	x_size, y_size,
	fun, image,
	elements)
	# return the output image
	return image

and here:

pyresample/pyresample/gradient/_gradient_search.pyx

Lines 240 to 248 in c930465

	one_step_gradient_search_no_gil(fake_data,
	src_x, src_y,
	xl, xp, yl, yp,
	dst_x, dst_y,
	x_size, y_size,
	indices_xy, indices,
	elements)
	return indices

My initial tests at naively adding with nogil: around these calls resulted in a segmentation fault. My hope is that someone like @mraspaud who is familiar with the algorithm can figure out why this might be happening.

[mraspaud]

For reference: I looked a this last week and was unable to pinpoint where the segfault was coming from. Cygdb wasn't very cooperative.