Readout error mitigation should calibrate on polynomial number of observables

Currently readout error mitigation in measure_observables calibrates each observables that it measures. While this is certainly valid, it scales badly -- each observable to be measured needs to be calibrated, so overhead is proportional to the number of observables to be measures (linear in the number of observables).

Instead one should characterize readout imperfections locally, i.e., calibrate how expectation of one body observables are distorted by readout, and extrapolate to many body observables. In this way the overhead would be independent of the number of observables to be measured, and depend only linearly on the number of qubits involved in the measurement.

This, of course, assumes readout errors are not correlated. But in the presence of two body correlations, one only needs to measure at most a quadratic number of observables. If N body correlations are significant, O(2^N) observables would be needed, but it is reasonable to expect that N is constant and independent of the number of qubits (and almost certainly N=2).

Start by reading the measure_observables implementation and its existing calibration path. Determine how calibration cost currently scales with the number of observables, then identify the tests or benchmarks covering readout error mitigation. Done should demonstrate local calibration with overhead that scales polynomially with the relevant correlations rather than with every measured observable.