Make C++ to Python state vector copying cheaper

[95-martin-orion]

Users have noted that calling qsim_simulator.simulate() from Cirq takes considerably longer than the qsim simulation itself. This is likely due to the added cost of copying results from C++ to Python, which in theory can be avoided.

To resolve this issue, the pybind layer should ensure that results from Python are captured in C++ objects without copying.

[sergeisakov]

Calling results = qsim_simulator.simulate() from Cirq takes about the same time as the qsim simulation itself. The results are not copied from C++ to Python. Calling results.state_vector() from Cirq may take considerable time for large circuits. The problem is in Cirq not in qsimcirq, see line 215 in https://github.com/quantumlib/Cirq/blob/master/cirq/sim/state_vector_simulator.py.

[95-martin-orion]

Aha! That means this is a victim of Cirq PR 3115, which is meant to prevent mutation of the state vector in other simulators. I'll redirect this issue to Cirq.