Benchmark-noise filters: compare against a hindsight oracle, not random drops

To show that a filter fixes a null A/B comparison (two runs of the same binary), comparing it with dropping the same number of benchmarks at random is too easy to beat: any rule that picks slow benchmarks wins by construction. A fairer reference is an oracle that drops the same number of benchmarks with the largest |log(B/A)|, which needs both runs, while the filter sees each run alone. In one case the random drops left a median 3.93% geomean error, the oracle 0.76%, and the filter 0.72%. The oracle is not a lower bound either, because it drops outliers in both directions; on a quiet pair it made the error worse (0.27% -> 0.82%). Also, when injecting load with multiprocessing children for ground truth, killing the parent leaves the children spinning into the next run; kill by parent PID (pkill -P).