Use bit shifts in generic and byte_shuffle reverse

[argilo]

Using bit shifts instead of bit fields seems to be much more compiler-friendly, and results in large speedups with both GCC & Clang, across x86_64, aarch64, and armhf. When compiling with Clang, the generic protokernel comes close to being the fastest protokernel on all three architectures.

Compiler	generic before	generic after	byte_shuffle before	byte_shuffle after
GCC x86_64	2687	389	3193	167
Clang x86_64	1398	46	4883	168
GCC aarch64 (Pi 5)	3406	1887	4007	458
Clang aarch64 (Pi 5)	1187	78	5089	111
GCC armhf (Pi 4)	6674	4331	5771	963
Clang armhf (Pi 4)	3322	321	6994	1022

[marcusmueller]

Obvious win! It's interesting how much clang and gcc diverge here, and also, that there's no consistency in which implementation wins. Fun fact, looking at disassebly, here: clang reworks the the generic bit shift implementation into the bintree_permute_top_donw code, basically, for AVX2 (which explains why run times are near identical).

[argilo]

Clang is also smart enough to use rbit on ARM, which explains why the generic protokernel does so well.

[jdemel]

LGTM. A nice performance win. Also, this is an interesting data point that Clang might produce faster code than GCC.