Validating LuaJIT optimizations about raptorjit

[lukego]

Here is a better visualization of the optimizations benchmark results based on the one in #53 (comment). Click to zoom:

[lukego]

Looking at this lovely picture I can spot two cases where disabling an optimization changed not only the average performance but also the shape of the distribution:

• roulette went from bimodel to consistent when loop optimization was disabled.
• nsieve went from consistent to bimodel when loop optimization was disabled.

The first one makes sense: loop optimization is more effective on some runs of roulette than others (that is the specific case that this benchmark exercises.) The second one surprises me: how come disabling loop optimization makes nsieve much slower and also splits the results into two clusters? This is likely worth understanding in order to identify another source of nondeterministic performance.