Benchmark CLI `compare` can have inaccurate results for some metrics

[popematt]

The compare command in the benchmark CLI always assumes lower is better. This is not true when measuring e.g. ops per second, and so the compare command can erroneously detect regressions for those metrics.