RDMA Feature Collection Support

@VoVAllen @Rhett-Ying

Quiver Team recently released our [RDMA based feature collection solution](https://github.com/quiver-team/quiver-feature) which is about 5x faster than `TensorPipe`'s solution. `quiver_feature.DistTensorPGAS` is a distributed tensor abstraction above heterogeneous memories and it can be easily integrated into DGL under pytorch ecosystem. We've also [released docs about how we use RDMA to get the best performance](https://github.com/quiver-team/quiver-feature/blob/main/docs/rdma_details.md) which might be helpful if you guys want to support RDMA feature collection using TensorPipe.