Use stb's paeth predictor

[kornelski]

The stb's implementation doesn't seem to be always faster.

For me on Apple M3 it's faster for 1, 6, and 8-byte pixels. About the same for 4. Half the speed for 2 and 3. I'm curious if that's universal across platforms. On a Ryzen most are faster, except 6 Bpp.

Bpp	1	2	3	4	6	8
Apple M3	y	N	N	=	y	y
Ryzen 5 5600X	y	y	y	y	N	y

I've used cargo +nightly bench -Fbenchmarks Paeth --bench unfilter -- --save-baseline=…

[fintelia]

Thanks for the measurements! I've incidentally also got a 5600X, so I can't offer any further data points

[HeroicKatora]

Thanks for the comparative measurements, can be closed with #539 as implemented and constrained to arch x86.