Break 256-bit vector int add/sub/mul into two 128-bit operations to avoid costly scalarization. Fixes PR10711.

llvm-svn: 138427