Add USE_CUDA macro in THD DataChannelGloo for non-GPU use case (#20186)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/20186
We're missing two USE_CUDA macro for GPU-related code in THD's DataChannelGloo.
Also adding GlooCache back to compilation.
Differential Revision: D15227502
fbshipit-source-id: f260e1cb294d662ba0c170931913b64287d62344
