add count_primes

[oscardssmith]

Based off of discussion in #99, this turns the existing prime sieve into a more efficient segmented one. This is about 5x slower than FastPrimeSieve (as tested on 10^10), but also only takes about 30 new lines of code. This will be a better foundation for #102

[haampie]

Notice that FastPrimeSieve.pcountprimes is threaded, which gives more speedup:

julia> @time Primes.count_primes(1, 10^10)
 22.230106 seconds (162.77 k allocations: 2.545 MiB)
455052511

julia> @time FastPrimeSieve.countprimes(10^10)
  2.389320 seconds (7 allocations: 277.109 KiB)
455052511

julia> @time FastPrimeSieve.pcountprimes(10^10)
  0.269437 seconds (14.94 k allocations: 4.173 MiB, 0.00% compilation time)
455052511

julia> @time FastPrimeSieve.pcountprimes(10^10, segment_length=128*1024)
  0.195094 seconds (243 allocations: 4.828 MiB)
455052511

So it's about a factor 114 slower on my 8 core / 16 threads AMD processor after a bit of tuning.

For reference:

$ primesieve 1e10
Sieve size = 256 KiB
Threads = 16
100%
Seconds: 0.179
Primes: 455052511

[oscardssmith]

I was comparing to single threaded since I also plan to eventually add multithreading to this code. I believe the main place improvement to be made is that currently, the inner loop has a divrem to calculate the index that accounts for a lot of the time in the algorithm, and I think it is removable.

[haampie]

Yeah, the offset can be saved after the last iteration of each prime number inside a segment

[oscardssmith]

Closing this since #102 is a better path forwards.