[AVX] Implement 256-bit vector lowering for SCALAR_TO_VECTOR.  This
largely completes support for 128-bit fallback lowering for code that
is not 256-bit ready.

llvm-svn: 125315