Added Machine-Prime, simplified nearest_prime

[JASory]

Added the low-memory variant of Machine-prime (using a modified BPSW test) as an optional dependency. The Criterion benchmark shows that it is approximately 16x faster than the existing version. Machine-Prime's other variants are faster still but require more memory and this is already a memory-intensive library.

MP is no-std and usable in const functions as well.

Simplified the next_prime and previous_prime functions

[JSorngard]

Very impressive!

The is_prime function gets a ~440% increase in throughput on my machine when the feature is on:

The plot of the execution time distributions is quite funny as well:

Red is with the feature off and blue is with it on.
Quite a bit faster.

[JASory]

The difference in benchmark is interesting. I consistently get 16x speed up. But that seems to be because I wrote/optimised it on lower-end hardware. I wonder if the trial division is somehow getting vectorised on your machine. Either way the worst-case is equivalent to only 2.5 fermat tests, so it'll beat it out pretty much regardless of what compiler optimisations are done.

On i5-5300u processor

time 12.021 ms -> 0.666 ms

Throughput 0.831M /s -> 15M/s

[JSorngard]

Interesting :o
Just using godbolt.org does not reveal any use of vector registers as far as I can tell from reading over the assembly, but modern processors are such massively complex beasts that there might be a technique I am not aware of that speeds it up a lot on my CPU.

My CPU is a 5800X3D, but I don't think the extra cache should help in this situation, it must be something else.

Regardless, I appreciate your excellent contribution!