[AArch64] Add tests for 128-bit shift variations.

It looks like there's still some room for optimization; the funnel shift
lowering is more efficient than the lowering of actual shifts.