Find interference-hit benchmarks from pyperf's runnable_threads

pyperf writes runnable_threads (from /proc/loadavg, Linux only) into every worker's run metadata, but nothing in pyperf acts on it. On a shared 8-vCPU VM, flagging benchmarks whose median runnable_threads exceeds cpu_count/2 (physical cores with SMT x2) found the damage from other tenants. Two runs of the same binary differed by 3.84% geomean; dropping the 10 flagged benchmarks (of 122) left 0.96%, while dropping 10 at random left a median of 3.95%. Any cutoff from 3 to 8 gave 0.9-1.6%. Rerunning only the flagged benchmarks is much cheaper than rerunning the suite. Remember that pyperf hoists values shared by all processes up to the benchmark-level metadata. On Windows the key is absent.