We were doing this late after nir_lower_io, but we can just reuse the core code. By doing it at this stage, we won't even set up the VS attributes as inputs, reducing our VPM size.