Careful microoptimisations for 2x performance improvement

[oxinabox]

This could do with some careful review.

It halves the time of the constructor (both for found, and not found),
and overall the tests suite runs 30% faster.

[oxinabox]

I think the final version might be 5x faster in some cases.

@xiaodaigh can you try this out on your task?
(String sorting was it?)

[oxinabox]

@xiaodaigh did you ever try this out?

[xiaodaigh]

Yep I tried just then. It can convert a 100m vector with 1m unique values into InternedStrings vector in about 100 seconds. I tried adding the same 100m vector again and it completed in 93 seconds. This seems quite slow compared to R's 13s.

I then tried to make an interned string pool with 1m unique strings and sample from that to get 100m strings. The conversion of 1m strings took 4.7s and the sampling took 120 seconds

using InternedStrings

@time samplespace = InternedString.("id".*dec.(1:1_000_000, 10));
srand(1);
@time svec = rand(samplespace, 100_000_000);

see R's equivalent

library(data.table)
pt = proc.time()
ss <- sprintf("id%010d",1:1e6)
system.time(x <- sample(ss, 1e8, replace=T))
print(x)
2+2
timetaken(pt)

[oxinabox]

On the v0.3.0 code (without micro opts)
I get

julia> using InternedStrings
julia> @time samplespace = InternedString.("id".*dec.(1:1_000_000, 10));
 10.471583 seconds (31.21 M allocations: 656.495 MiB, 26.22% gc time)

julia> srand(1);
julia> @time svec = rand(samplespace, 100_000_000);
322.384380 seconds (799.95 M allocations: 14.156 GiB, 56.60% gc time)

On the 0.4.0 code, with the microopts I get:

julia> using InternedStrings
julia> @time samplespace = InternedString.("id".*dec.(1:1_000_000, 10));
  6.838613 seconds (13.23 M allocations: 351.488 MiB, 11.40% gc time)
julia> srand(1);
julia> @time svec = rand(samplespace, 100_000_000);
 90.561751 seconds (100.00 M allocations: 2.235 GiB, 16.08% gc time)

I don't see how the R example is the same at all though.
That isn't interning strings. Is it?
My R is pretty weak

To me that looks like

julia> @time samplespace = String.("id".*dec.(1:1_000_000, 10));
  0.231330 seconds (3.01 M allocations: 130.061 MiB)

julia> srand(1);
julia> @time svec = rand(samplespace, 100_000_000);
 22.346334 seconds (2.18 k allocations: 763.058 MiB, 75.21% gc time)

As an aside these tests are kinda worse case scenario for julia dodgy since they are using nonconst globals.