Shared cloud VM: a whole pyperformance run can shift ~2% with no metadata trace

On a shared 8-vCPU cloud VM, three full pyperformance runs of the same CPython binary in the same JIT-on configuration: the first two agreed within 0.27% geomean, and the third, a day later, was about 2% faster on most benchmarks (61-78% faster, median ratio 0.98-0.99). Per-worker runnable_threads was lower during the third run, but regressing the per-benchmark log ratio on the change in mean runnable_threads still leaves -1.5 to -1.8%, and pyperf's recorded mean cpu_freq differed by only 0.4%. Filters that drop burst-hit benchmarks can't remove a shift like this. Treat whole suite runs, spread over time, as the replicates (several per side, interleaved). Agreement between two back-to-back runs doesn't prove the absolute level is right.