Locality-Aware Work Stealing on Multi-CPU and Multi-GPU Architectures