We had to revert #1661 due to major performance issues on Add/Inc/Observe methods for cumulatives #1748

While #1661 added benchmarks, it seems they were unrealistic (e.g. around context switching and extra work in between). We also didn't have a good judgement on that 10ms overhead. We need to make sure:

1. There are benchmarks we can rely on.
2. Our release process (or even per PR) ensures we run those and detect potential regressions.

Help wanted!