[X86][AVX] Combine non-lane crossing binary shuffles using X86ISD::VPERMV3

Some of the combines might be further improved if we lower more shuffles with X86ISD::VPERMV3 directly, instead of waiting to combine the results.

llvm-svn: 359400