Thin out SIMD libraries into common code

[solidpixel]

We should aim to minimize the amount of code in the per-ISA SIMD headers, moving things we can into a common header. We've done this for the 4-wide code, although it could do with a cleanup pass, but we should do it for the 8-wide ISAs too.

[solidpixel]

Removals:

• Remove vint4 hadd_s() #527 (unused)