Stop returning values in entry points and improve predict server tests

[jasonrute]

Change 1: No longer return values in g2t-tfgnn-train, g2t-server, and g2t-hmodel-train. These would be treated as errors and cluttered output. But the values can still be used for tests.

Change 2: Change the format of the predict server expected outputs. There were two problems:

• In our tests, there are many outputs with the same confidence (this is because of symmetry in local arguments). If we change the algorithm to select the best tactics (like using beam search), then the order could change. This standardizes it for the test.
• By making the predict server expected outputs a list, it works a bit better when the test fails. So now, it gives a (more) readable error, like:

  E         At index 8 diff: {'_type': 'TacticPredictionsGraph', 'contents': {'predictions': []}} != {'_type': 'TacticPredictionsGraph', 'contents': {'predictions': [{'arguments': ['node-1-0-propchain_dep.start'], 'confidence': 0.5 ± 5.0e-05, 'ident': 5294290463909144854}, {'arguments': ['node-1-2-propchain_dep.th0'], 'confidence': 0.5 ± 5.0e-05, 'ident': 5294290463909144854}]}}

  (Showing in this case that the model returned an empty prediction list for one of the predictions, but we were expecting two predictions.)

[mirefek]

Is the hmodel-predict-server supposed to work now? It still doesn't for me but you changed some of the relevant files (so I am not sure what the change is about).

[jasonrute]

@mirefek Sorry, I should have made clear, I manually changed all the expected-predict-server.yaml files with a script. That way I could be sure the results are the same except for the permutation of the order the results are listed in (if there is a tie in the confidence). That is why the hmodel server expected results file changed even though the test is still broken.

[mirefek]

I see