Support reference sample in simulators

[rafaelha]

To mirror stim, the compilers should have an argument skip_reference_sample: bool = False.

In our case, skip_reference_sample: bool = False actually takes extra work, since we have to compute a noiseless sample. However, we should still support this and make it the default to be compatible with stim.

From the stim docs:

def compile_sampler(self, *, skip_reference_sample: bool = False, seed: object = None) -> stim.CompiledMeasurementSampler:
    """Returns a CompiledMeasurementSampler, which can quickly batch sample measurements, for the circuit.

    Args:
        skip_reference_sample: Defaults to False. When set to True, the reference sample used by the sampler is
            initialized to all-zeroes instead of being collected from the circuit. This means that the results
            returned by the sampler are actually whether or not each measurement was *flipped*, instead of true
            measurement results.

            Forcing an all-zero reference sample is useful when you are only interested in error propagation and
            don't want to have to deal with the fact that some measurements want to be On when no errors occur.
            It is also useful when you know for sure that the all-zero result is actually a possible result from
            the circuit (under noiseless execution), meaning it is a valid reference sample as good as any
            other. Computing the reference sample is the most time consuming and memory intensive part of
            simulating the circuit, so promising that the simulator can safely skip that step is an effective
            optimization.

[rafaelha]

After discussing with @ChenZhao44: A reference sample only makes sense if detectors/ observables are deterministic. We can (often, but not always?) read off this property directly from the reduced ZX graph.

If we want to support skip_reference_sample, an error needs to be raised for non-deterministic detectors/observables.

It is also conceivable to have a reference sample only for the detectors. This was what handled a case where detectors are deterministic but the observable is not.

This is a low priority issue.