RFC: temp/musl: memcpy: use THUMB optimizations

It's not polished pull request, but please consider merging this until this get (probably in different form) upstreamed.

The code itself should be absolutely OK.

Also I tested, results of benchmarks (approx 40% additional memcpy performance) are in commit.

I also opened Alpine MR: https://gitlab.alpinelinux.org/alpine/aports/merge_requests/3130