Implement scripts for detection of performance regressions across CUB versions

Tuning infrastructure is currently used to find better tuning variants. Reusing it for the detection of performance regressions is preferable since there should be a single source of truth for workloads and benchmarks we care about. The detection of performance regressions should be based on sample distribution analysis. We can start with Mann–Whitney U test, and if it's likely that two distributions do not belong to a single one, proceed with the ratio function based on the Harrell-Davis quantile estimator. It might also be worth it to experiment with outlier detection.