Missing documentation about the environment variables used by distributed training

## 📚 Documentation

There are a bunch of variables used by distributed training but are not documented anywhere:

* `'DGL_IP_CONFIG'`
* `'DGL_NUM_SERVER'`
* `'DGL_NUM_CLIENT'`
* `'DGL_CONF_PATH'`
* `'DGL_GRAPH_FORMAT'`

Currently they are set by the `launch.py` script implicitly and used by `dgl.distributed.initialize`.

https://github.com/dmlc/dgl/blob/8292bf329fbb2418aa6d10f140b0a8150c090d3c/python/dgl/distributed/dist_context.py#L176-L218

Two considerations:
* Passing values by environment variables is quite implicit. Passing an explicit configuration object may be more preferred.
* `dgl.distributed.initialize` should clearly document these variables because they change the behavior of how processes are launched.

@zheng-da What do you think?

Related: #4272