Transpile parallel experiment sub circuits

[chriseclectic]

Summary

Parallel Experiment will now transpile sub experiments according to their individual transpile options before combining them. The combined circuits will then be transpiled again according to any options set in the parallel experiment transpile options.

Details and comments

This is applying the same logic from #667 for batch experiment to the more complicated case of parallel experiments.
This will require slight modification if #677 is merged first, or vice versa.

[yaelbh]

I think it's important to add tests that specifically check this new behavior. Especially as a safety net, for the case that someone modifies the code.

[chriseclectic]

@yaelbh I'm going to wait until #667 and #677 are merged before making any further changes to this PR.

[nkanazawa1989]

This is high-level comments. Overall I think 2 stage transpile is reasonable direction especially for this class since we could run into the situation where different kind of experiment instances with different run options are combined.

I'm also interested in validation of parallel circuit when component experiment contains no-adjacent qubit pairs, i.e. first transpile may inject swaps and this may conflict with another combined instance. In this case parallel experiment will be no longer helper function, indeed it may result in unexpected outcome, since this combine logic doesn't take care of equivalence of execution circuit, i.e. if there is swap circuit ahead of the experiment circuit, that will start from non-zero init state. We need to write validator, or, just limit coupling map of component experiment, i.e. base experiment.

[nkanazawa1989]

This remind me the discussion about coupling map for QV. Sometimes transpiler couples two qubits with swapping via ancillary qubits, this will cause a problem, for example, when sub_circ is mapped to the non-adjacent qubits. The simplest solution to this would be automatically generate valid coupling map in the base experiments not to allow such experiment configurations, i.e. ancillary qubits should be explicitly defined in constructor argument qubits. If you don't limit the coupling map, please implement the logic and add the test for this edge case.

class EntangleExperiment(BaseExperiment):
    def __init__(qubits, backend):
        super().__init__(qubits, backend=backend):

    def circuits(self):
        qc = QuantumCircuit(2)
        qc.cx(*self.qubits)
        return [qc]

backend = SomeFakeBackend()    # this should be some Falcon 27Q device
exp1 = EntangleExperiment(qubits=[0, 4], backend=backend)
exp2 = EntangleExperiment(qubits=[2, 5], backend=backend)    # if you set 1, 2 instead you'll run into another edge case

pexp = ParallelExperiment([exp1, exp2])
pexp.circuits()   # do some validation for this.

[chriseclectic]

This is independent of parallel experiment and this sort of validation should happen in base experiment for each individual component experiment.

I just tried testing something towards this but it was much more difficult than expected and ran into a lot of unexpected transpiler errors with the coupling pass, so I think this needs to be done later. The general issue is Terra does not like truncated coupling maps, it will raise a bunch of transpiler errors about CouplingError: 'coupling graph not connected'