Fix an issue where fast math flags were dropped during scalarization.

Most portions of InstCombine properly propagate fast math flags, but
apparently the vector scalarization section was overlooked.

llvm-svn: 262376