Improve speed of random_agent

[Tortar]

It was actually very slow before, consider this silly model I was using to benchmark a different thing:

using Agents, Random, BenchmarkTools

function init_model()
	model = StandardABM(GridAgent{2}, GridSpace((50, 50)), rng=Xoshiro(42))
	for _ in 1:1000
		add_agent!(model)
	end
	return model
end

function model_step!(model)
	add_agent!(model)
	remove_agent!(random_agent(model), model)
end

model = init_model()
@benchmark step!(model, dummystep, model_step!, 10^5)

before:

julia> @benchmark step!(model, dummystep, model_step!, 10^5)
BenchmarkTools.Trial: 27 samples with 1 evaluation.
 Range (min … max):  183.797 ms … 194.004 ms  ┊ GC (min … max): 0.00% … 0.00%
 Time  (median):     187.722 ms               ┊ GC (median):    0.00%
 Time  (mean ± σ):   188.308 ms ±   2.707 ms  ┊ GC (mean ± σ):  0.13% ± 0.37%

                 ▃ █     ▃
  ▇▁▁▇▁▁▁▁▁▇▁▇▁▁▇█▁█▇▁▇▇▁█▁▇▁▁▁▁▇▇▁▇▇▇▁▇▁▁▁▁▇▁▁▁▁▁▁▁▁▇▁▁▁▁▇▇▁▁▇ ▁
  184 ms           Histogram: frequency by time          194 ms <

 Memory estimate: 5.19 MiB, allocs estimate: 100169.

after:

julia> @benchmark step!(model, dummystep, model_step!, 10^5)
BenchmarkTools.Trial: 306 samples with 1 evaluation.
 Range (min … max):  15.251 ms … 26.520 ms  ┊ GC (min … max): 0.00% … 0.00%
 Time  (median):     15.822 ms              ┊ GC (median):    0.00%
 Time  (mean ± σ):   16.360 ms ±  1.440 ms  ┊ GC (mean ± σ):  1.50% ± 3.67%

    ▄█▆▁
  ▃▅████▇▆▃▃▄▅▄▄▃▃▁▃▃▃▃▄▃▃▃▃▃▂▂▂▃▂▃▁▁▂▁▁▁▁▁▃▂▁▁▁▁▂▁▂▁▁▁▁▁▂▂▁▂ ▃
  15.3 ms         Histogram: frequency by time        21.7 ms <

 Memory estimate: 5.19 MiB, allocs estimate: 100162.

A micro-benchmark tells that now random_agent for a StandardABM is 40x faster!

I exported a new function random_id in the process, which is in line to the other random functions which export also the id version

[Tortar]

I have to say that this means that probably there is something to improve in Base, since that the previous method is so slow sounds strange

[codecov-commenter]

Codecov Report

Merging #897 (737312d) into main (a71cc12) will increase coverage by 12.09%.
The diff coverage is 60.00%.

❗ Current head 737312d differs from pull request most recent head 90a22d6. Consider uploading reports for the commit 90a22d6 to get more accurate results

@@             Coverage Diff             @@
##             main     #897       +/-   ##
===========================================
+ Coverage   80.27%   92.37%   +12.09%     
===========================================
  Files          44       32       -12     
  Lines        3027     2334      -693     
===========================================
- Hits         2430     2156      -274     
+ Misses        597      178      -419     

Files	Coverage Δ
src/core/model_abstract.jl	89.06% <66.66%> (-0.03%)	⬇️
src/core/model_concrete.jl	88.23% <50.00%> (-1.18%)	⬇️

... and 18 files with indirect coverage changes

📣 We’re building smart automated test selection to slash your CI/CD build times. Learn more

[Tortar]

Actually it was a O(N) operation while this one is not, this means that the improvement in speed is arbitrary large!

[Datseris]

I don't think it is a bug in base. It is probably an inferrence issue since the agent container is complicated. The function barrier makes the inference work.

DOn't forge tto add changelog entry and then merge!

[Tortar]

I think it should be JuliaLang/julia#51605 , let's see :-)