Fix handling of circuit metadata

[hhorii]

Summary

#1435

Details and comments

to_json is called for each circuit header.
A header may include metadata and metadata can be a python object.
This PR changes serialization of circuit headers to use py::handle without serializing to a json.

Another solution can be in _assemble() to remove all unnecessary metadata from input.

[jakelishman]

I'm not very knowledgeable about the C++/Python handoff with pybind11, so a slight question: what we need out of the header seems very simple - can we just leave the metadata in Python space, and reattach it to the new output circuit (if there is one)? It feels like we could save time by just not serialising it, if we don't use it.

[hhorii]

what we need out of the header seems very simple - can we just leave the metadata in Python space, and reattach it to the new output circuit (if there is one)?

I believe that py::handle can be passed through C++ space. I will try it.
The current implementation of ExperimentResult uses json_t. To pass py::handle instead of json_t, as other codes, we need to make [ExperimentResult] class a template class to handle both of json_t and py::handle, and modify to_python(ExperimentResult) to handle the both types.

[kevinsung]

Please add a test that ensures the repro code from #1435 works.

[hhorii]

Please add a test that ensures the repro code from #1435 works.

I added test cases to check metadata is copied correctly.

[hhorii]

TODO:

• Use pop instead of del
• Store an index of original circuit for parameterized circuit.
• Change _assemble instead of _run
• Do not change codes to copy headers in controller

[hhorii]

Aer does not use Metadata to simulate circuits. Therefore, I believe, we can reduce metadata from qobj before simulation.

Previously, metadata of QuantumCircuit is copied to QobjExperimentHeader in a qobj. ThisQobjExperimentHeader is returned as header of ExperimentResult. Whit this change, header of ExperimentResult does not keep metadata of QuantumCircuit because it is removed before generating qobj.

[kevinsung]

LGTM. @chriseclectic or someone else with write access please take a look.

[mtreinish]

Overall LGTM, just one question inline and a suggestion on the release note. Thanks for updating this

[hhorii]

I realized that id in metadata is necessary for clustering. QObj is divided to sub-QObjs (first level) and each sub-QObj can be divided to subsub-QObj (second level). If id is removed in the first level, there is no way to merge results in second level.