Cannot set NVCC FLAGS by environment variable

[brooks0517]

cannot pass nvcc(cuda) flags to gloo by environment variable, such as '-D_GLIBCXX_USE_CXX11_ABI=0'.
It's not convenient when compile gloo as third party, such as when compiling PyTorch.

[yichi-yang]

It seems this line blocks propagation of CMAKE_{C,CXX}_FLAGS (e.g. -D_GLIBCXX_USE_CXX11_ABI=0). You can try commenting it out to get it to compile with PyTorch. There is a related issue in the PyTorch repo. It will be helpful if someone can point out on which platform do we need the set(CUDA_PROPAGATE_HOST_FLAGS OFF) line?