nir: Do not reset shared_size in nir_lower_io.

I'd like to use raw shared intrinsics already for some raytracing stuff before this pass gets called and this was a real pitfall.

This mirrors scratch_size and constant_data_size.