Improve performance of average snapshots

[chriseclectic]

Summary

Previously the mean and variance were returned as a copy of stored
accumulated data. This allows them to modify the stored data and return
a reference that can be moved into result containers which can help
avoid copies for large data objects like probability and density matrix
snapshots. To make this work it was also necessary to allow JSON conversion of
results to be non const.

Details and comments

Memory profiler comparisons for following circuit:

nq = 14
circ = QuantumCircuit(nq)
circ.snapshot_density_matrix('state')

backend = QasmSimulator()
backend_options = {
    "method": "density_matrix",
    "fusion_enabled": False,
    "truncate_enabled": False,
    "max_parallel_threads": 1
}
qobj = assemble(circ, shots=1)

start = time.time()
result = backend.run(qobj, backend_options=backend_options).result()
stop = time.time()
time_py = stop - start
time_cpp = result.metadata['time_taken']

For this PR: time_py = 29.2 s, time_cpp = 26.7 s
For master: time_py = 31.8 s, time_cpp = 58.3 s

And memory profiling PR:

vs Master:

[vvilpas]

If we move the mean and variance results, subsequent calls to those functions will yield wrong results. Probably we don't need to call them again but if that's the case we may need to implement something to recompute the values, as normalize won't be executed again as normalized_ will be true. What do you think?

[chriseclectic]

It should be fine to call mean and variance multiple times, only the first call will result in normalizing the data (this is because normalize does nothing if normalized_ = true). But if you move them then yes, you've destroyed the container. I think this is like any other C++ object, after you used move you shouldn't use it again. If you want to use it again you should copy or reference the data rather than move, however i don't think this situation comes up in our current use case.

[vvilpas]

I think the movement is a bit more hidden because we are moving specific members instead of the whole object, but as we're not gonna reuse them it seems we are ok.

[vvilpas]

Lgtm! Nice performance improvement!