fixed TargetArchs implementation in case when HAVE_CUDA=false, added initial structure for multi_gpu sample