Support worker/server nodes use cluster ip

[solin319]

Use 0.0.0.0 to bind all nodes' listening port.
This can make all nodes use cluster ip in kubernetes.
We can set DMLC_NODE_HOST and DMLC_PS_ROOT_URI with cluster ip when use kubernetes to launch distribute jobs.
@mli

[sswv]

In my Kubernetes based MXNet cluster, I also found the issue of using Cluster IP. Cluster IP can be accessed through virtual routing but it cannot be bound by socket. Thus, it is not good to use DMLC_NODE_HOST for both binding at server and accessing at client. I had some tricked modification in my MXNet cluster to solve it.

I think it is necessary to differentiate "the IP for socket binding at server" and "the IP for accessing at client". It is a good idea to use 0.0.0.0 simply when binding socket at server. TensorFlow also did that in its GrpcServer.

[coldsheephot]

I think this is necessary when I use k8s to create mxnet distributed job.