[x86] Teach the new vector shuffle lowering to use VPERMILPD for

single-input shuffles with doubles. This allows them to fold memory
operands into the shuffle, etc. This is just the analog to the v4f32
case in my prior commit.

llvm-svn: 218193