make the fallback implementation of slidehash vectorize better

[folkertdev]

I'm picking 16 because on SSE the loop is just unrolled, and on other targets with 256-bit wide registers they could actually use the full width.

The table is always a size that is a power of 2, either 1 << 16 for state.head or between 1 << 8 and 1 << 15 for state.prev (it's based on the window size). So it's totally legit to use chunks here, and there is no risk of ignoring elements.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Files with missing lines	Coverage Δ
zlib-rs/src/deflate/slide_hash.rs	99.14% <100.00%> (+0.03%)	⬆️

[folkertdev]

On neon our custom version appears to generate better code (hard to judge, but it's fewer instructions in the hot loop)

https://godbolt.org/z/sd3Me8fnG

[folkertdev]

turns out a chunk size of 32 generates less code because that appears to be, on aarch64 and x86_64, how much LLVM wants to unroll (probably based on the number of concurrent loads that the CPU can do).