Use qsim for repeated sampling

[95-martin-orion]

Fixes #503.

*_sample and *_sample_final need to be different functions because they have incompatible behaviors:

• *_sample (the existing methods) sample from measurements in a single execution of a circuit.
• *_sample_final (the new methods) sample repeatedly from the final state of a circuit.

[sergeisakov]

LGTM. Does it make sense to add some non-trivial test?

[95-martin-orion]

LGTM. Does it make sense to add some non-trivial test?

I considered adding tests for #502, but a 33-qubit circuit of any depth would be excessive for a unit test. Other than that, the existing Python tests cover repeated sampling pretty thoroughly - is there anything extra you see that needs testing?

[sergeisakov]

LGTM. Does it make sense to add some non-trivial test?

I considered adding tests for #502, but a 33-qubit circuit of any depth would be excessive for a unit test. Other than that, the existing Python tests cover repeated sampling pretty thoroughly - is there anything extra you see that needs testing?

It seems that the tests are sort of trivial. For instance, the circuit in test_qsim_run_vs_cirq_run (line 405 in qsimcirq_test.py) just swaps the |0> amplitude and the run method always measures the same bit string. I was thinking of, say, a 4- or 5-qubit circuit with less trivial output and running a few thousands repetitions and comparing the measurement histogram with the actual probabilities. I am not sure if it makes sense to add such a test.

[95-martin-orion]

The pybind call is simple enough that I think it's safe to rely on the C++ tests for verifying sampling behavior. The only new complexity on the Python side is in the bit-management of converting raw_results to full_results - I don't think a circuit with multiple possible sampling outcomes adds much coverage in this regard.

For reference, the existing test_qsim_run_vs_cirq_run uses a single-outcome circuit to specifically examine the behavior of that outcome. The result is asymmetric (0101) to check for bit-ordering errors, and multiple measurements are extracted to verify that mapping from raw data to keyed results is working.

[sergeisakov]

SGTM