cleanup _mm256_permute2x128_si256_0x21

[Validark]

This happens to help the compiler directly emit vperm2i128 ymm1, ymm1, ymm3, 0x21.

Also made the code a bit more resilient to change by basing the conversion on the vector bit size divided by 4 rather than if (chunk_len == 32).

[travisstaloch]

This looks good to me. I'm going to merge but @sharpobject let us know if you have any objections as this overwrites some of your changes. Thanks! :+1:

[sharpobject]

Seems great!