Improve performance of color distance calculations by kernel fusion

[grlee77]

Description

This MR is purely to fuse many separate kernels into a single elementwise kernel for each of the "color distance" functions. This is expected to result in substantial performance improvement by reduced kernel launch overhead and because a single pass through the image is much more memory efficient than many separate kernel calls.

There is not expected to be any change in behavior (existing tests must continue to pass)

Benchmark Results

I added benchmarks for these functions and compared the results before and after this change.

The acceleration values are the relative speedup as compared to the scikit-image implementation

For a pair of 32-bit LAB images of shape: (512, 512, 3)

function	acceleration (old)	acceleration (new)
deltaE_cie76	5.49	21.85
deltaE_ciede94	37.60	65.09
deltaE_ciede2000	12.54	109.71
deltaE_cmc	27.18	89.23

For a pair of 32-bit LAB images of shape: (3840, 2160, 3)

function	acceleration (old)	acceleration (new)
deltaE_cie76	31.12	250.74
deltaE_ciede94	79.15	117.50
deltaE_ciede2000	71.36	132.73
deltaE_cmc	70.65	154.83

[gigony]

Thanks @grlee77 !
I don't understand the internal algorithm but this looks good to me!

[grlee77]

Not sure that there is a common use case for actually running these on full images or if it is most commonly used just to compare the perceptual difference of single pairs of colors. If the later, then the CPU will be better for running that case.

I think perhaps there could be use cases like comparing perceptual differences among all pairs of colors in a categorical colormap, etc.

[grlee77]

/merge