Make HMAC CPU formats use SIMD and OpenMP fully

[solardiz]

I was surprised to find that our HMAC CPU formats use scalar hashing code in set_key even when building with SIMD. This also puts such code out of OpenMP parallel sections. We should probably replace this with equivalent functionality inside crypt_all, reached when new_keys is set (we already have some code only reached on new_keys in there, but we could move more code to there).

No wonder there's such a big difference between Many vs. Only one salt speeds for these formats. These changes should improve the Only one salt speed (as well as actual cracking with few different salts).