In regions destined for GPU offload, computing an address_of means
getting device address directly - no need (and actually incorrect) to
insert a runtime call to get the address. This was already working for
regions such as `gpu.launch` - but now it applies to acc regions as
well.