multi_CUDA with unified memory

[pdziekan]

If thrust vectors with Unified Memory allocators were used, we would not need explicit copying between GPUs. See:
https://devtalk.nvidia.com/default/topic/987577/-thrust-is-there-a-managed_vector-with-unified-memory-do-we-still-only-have-device_vector-cuda-thrust-managed-vectors-/?offset=4
https://devtalk.nvidia.com/default/topic/830260/using-thrust-to-sort-unified-memory-buffer-/
http://sbel.wisc.edu/documents/TR-2014-08.pdf