[X86] LowerABD - improve pre-SSE41 handling for v16i8/v4i32 nodes

The generic expansion still causes a problem for SSE targets without BLENDV/select node, but we can create a custom lowering until that can be addressed.