[Feature] Max readout and consecutive labeling for networkx

[mufeili]

Description

Checklist

Please feel free to remove inapplicable items for your PR.

• The PR title starts with [$CATEGORY] (such as [Model], [Doc], [Feature]])
• Changes are complete (i.e. I finished coding on this PR)
• All changes have test coverage
• Code is well-documented
• To the my best knowledge, examples are either not affected by this change,
  or have been fixed to be compatible with this change
• Related issue is referred in this PR

Changes

• As requested in How can I implement the function dgl.max_nodes? #306 , a naive implementation for max readout on nodes and edges is provided with max_nodes and max_edges in batched_graph.py. A more efficient implementation will require segment_max. The test cases are provided in tests/pytorch/test_readout.py
• As discussed in Creating DGLGraph from non-consecutive int nodes #334 , if we want to create a DGLGraph object from a networkx graph with non-consecutive node labels, we relabel the nodes using consecutive integers first before converting it into a DGLGraph. The changes are from_networkx in graph.py, graph_index.py, immutable_graph_index.py. The test cases are provided in tests/compute/test_basics.py.
• In from_networkx of graph.py, I also convert the networkx graph to be directed first, otherwise only half of the DGLGraph edges will have features.

[yzh119]

@mufeili does the segment_max you mentioned similar to segment_sum in tensorflow?

[mufeili]

@yzh119 Yeah, see doc and code. @HQ01 told me he's interested in doing it.

[yzh119]

i see, i've read the code before

[BarclayII]

I think later on we will need to replace the for loop in max_nodes etc with a custom op, e.g. scatter_max.

[mufeili]

@BarclayII I agree, this is just a very naive workaround.

[jermainewang]

LGTM. Some minor comments.

[jermainewang]

Great! Thanks Murphy.