Merge: NSieve optimization
The `shootout_nsieve` benchmark has been present in tests for a long time (longer than I can remember actually).
Up until now, its performances have been constant and quite good, since the work was done in less than 15ms.
However, due to its use of `FlatBuffer` and indexed access, it degenerates when using UTF-8 (see #1277) and the run time of `shootout_nsieve` scales from `0m0.011s` to `1m59.870s`.
This PR fixes that.
Pull-Request: #1583
Reviewed-by: Jean Privat <jean@pryen.org>