[X86] Force shuffle lowering to only create X86ISD::VPERM2X128 with 64-bit...

[X86] Force shuffle lowering to only create X86ISD::VPERM2X128 with 64-bit element types so we can remove some patterns from isel.

Intrinsic handling is still creating these nodes with 32-bit elements as well. But at least this gets rid of 8 and 16.

Ideally, someday we'll convert the intrinsics to generic vector shuffles and remove the intrinsics.

llvm-svn: 312702