DevicePacker doesn't pack current data after swap when using cudaGraph API

[cwpearson]

stencil/src/packer.cu

Lines 150 to 158 in 1a4c722

	void DevicePacker::pack() {
	assert(size_);
	#ifdef STENCIL_USE_CUDA_GRAPH
	CUDA_RUNTIME(rt::time(cudaSetDevice, domain_->gpu()));
	CUDA_RUNTIME(rt::time(cudaGraphLaunch, instance_, stream_));
	#else
	launch_pack_kernels();
	#endif
	}

When using CUDA graph, the kernels are recorded with whatever the "current" gridpoint data is.
After a swap, that is wrong, but the replayed kernels will still use it.