Parallelization of sub-dnn processing

[dan-dnn]

I have updated the following:

-Parallelize sub-dnn among workers
-Reduce unnecessary 'cudaDeviceSynchronize()'
-Change to 'cudaMemcpyAsync'
-Support CMake

[angry-crab]

Thanks for the PR. Regarding nvsci_headers, according to NV, it could be downloaded from L4T BSP source. https://forums.developer.nvidia.com/t/where-are-the-nvsci-headers/221091

As we discussed, I'll add this information to README.

Would you be able to remove nvsci_headers from the PR? As for CMakeList, I guess one way we could do is to include the path as usual. Please let me know your thoughts

include_directories (${gflags_INCLUDE_DIR} ./include /PATH_TO_NV_HEADERS/include/nvsci_headers)