[x86] Make the single-input v8i16 lowering directly recurse rather than

going back through the entire vector shuffle lowering.

This is an important step to being able to re-use this logic.

llvm-svn: 230743