Inline `Atomic` and `Ref` ops and unroll This should give more accurate results such that it should be possible to match results with other more accurate benchmark results such as a benchmark of an optimized lock using `fetch_and_add`.