[imurashka]

Two small changes in encoder.rs, the output bytes stay the same:

• get_block: when there is no subsampling (both strides are 1) it copies each row of 8 pixels from a slice, without the index math and bounds check per pixel.
• quantize_block: quantizes in natural order first and then reorders to zig-zag, so the quantize loop can be vectorized.

The criterion encode bench on an Apple M4 Pro, main vs this branch:

bench	main	this branch	change
encode rgb 100	46.30 ms	44.68 ms	-3.5%
encode rgb 4x1	24.64 ms	23.88 ms	-2.9%
encode rgb progressive	25.38 ms	24.57 ms	-3.1%
encode rgb optimized	115.0 ms	99.8 ms	-13.2%
encode rgb optimized progressive	105.2 ms	101.0 ms	-4.0%
encode rgb mixed	202.8 ms	194.4 ms	-4.1%

With 4:4:4 sampling the gain is bigger, because every plane takes the new path: on x86-64 a 1200x630 RGBA image at quality 82 went from 6.29 ms to 5.57 ms. I checked the output is byte-identical there, with 4:4:4 and 4:2:0. cargo test, cargo fmt --check and clippy pass.