Extract should be able to skip the input tensor names

[zhenhuaw-me]

We should be able to walk the graph starting with the output tensor names and auto infer the input names if not given.

It would be interesting to figure out if the user provided input tensor names and output tensor names don't cut a subgraph.