Utilize the stronger guarantees on the call arguments from the
harvester, which now always calls hwrngs with the buffer length
multiple of the word size. This allows to remove the excessive memory
accesses to temporary buffer when saving the entropy word.

Streamline the assembly and unify it between i386 and amd64.

Reviewed by: markm, des
Approved by: so (des)
Sponsored by: The FreeBSD Foundation
MFC after: 2 weeks