portable_simd version of Avg (4bpp)

[Sentimentron]

Implements a RGBA version of the Avg filter with portable_simd intrinsics.

CPU	Baseline	Result	Speedup
Arm Cortex A520	415.9 MiB/s	707.4 MiB/s	70.08%
Arm Cortex X4	2053.9 MiB/s	2334.9 MiB/s	13.68%
Apple Silicon M2	2053.9 MiB/s	2173.5 MiB/s	3.62%
AMD EPYC 7B13	2425.8 MiB/s	2150.8 MiB/s	-11.34%

Marked as draft until #632 is completed.

[Sentimentron]

AI disclosure: I wrote a original sliding-window portable_simd implementation of the Paeth filter (3bpp) and optimized it for best performance on the Cortex A520. I then used the Gemini family of LLMs provided by my employer to automatically adapt this code to the Avg filter from a written description, then optimize it to achieve the best possible code-generation and performance across all other micro-architectures in simulation. This PR is derived from that output, but includes documentation and other cleanups.

[okaneco]

There's another 4bpp case for the first row, where previous.is_empty(), not sure if you've tried that already.

image-png/src/filter.rs

Lines 612 to 624 in f33b850

	BytesPerPixel::Four => {
	let mut prev = [0; 4];
	for chunk in current.chunks_exact_mut(4) {
	let new_chunk = [
	chunk[0].wrapping_add(prev[0] / 2),
	chunk[1].wrapping_add(prev[1] / 2),
	chunk[2].wrapping_add(prev[2] / 2),
	chunk[3].wrapping_add(prev[3] / 2),
	];
	*TryInto::<&mut [u8; 4]>::try_into(chunk).unwrap() = new_chunk;
	prev = new_chunk;
	}
	}

[Sentimentron]

I hadn't tried it - wrote some quick code for it but it seems that the unfilter benchmark doesn't test this edge case... 🤔

[Sentimentron]

Also, if any contributors have access to some Intel hardware, could they give this portable_simd version a try? (Otherwise I'll cfg-gate it off in a subsequent version to avoid the AMD Epyc 7B13 regression).