use a stable rng from StableRNGs for reproducibility

[t-bltg]

Is is currently very hard (impossible ?) to rely on Graphs across different julia versions: this is observed in Plots.jl ci through downstream test of GraphRecipes.jl.

The reason is that the RNG across julia minors is documented to be unreliable in terms of reproducibility.

This makes ci testing via image comparison difficult unless duplicating reference files for different julia versions which is quite cumbersome.

It is advised from julia docs, to use StableRNGs.jl instead.

PR:

1. use StableRNGs in tests;
2. fix methods signatures using random related calls, to take a rng and seed;
3. update ci workflows.

[codecov]

Codecov Report

Merging #170 (0028764) into master (a3fb98b) will decrease coverage by 0.07%.
The diff coverage is 93.85%.

@@            Coverage Diff             @@
##           master     #170      +/-   ##
==========================================
- Coverage   97.54%   97.47%   -0.08%     
==========================================
  Files         109      109              
  Lines        6314     6332      +18     
==========================================
+ Hits         6159     6172      +13     
- Misses        155      160       +5     

[simonschoelly]

Thanks for this PR, I think StableRNG might indeed make sense for testing. But this PR would also add it outside of tests? I think we should avoid that, and only use StableRNG for the tests.

We currently might still have some random function that take only a seed instead of a random generator. These functions should be refactored so that they use rng_from_rng_or_seed as a temporary workaround - we then might get rid of the seed in the next Major release.

[t-bltg]

I think we should avoid that, and only use StableRNG for the tests.

I've restored the GLOBAL_RNG and used StableRNGs only in the tests.

These functions should be refactored so that they use rng_from_rng_or_seed as a temporary workaround

I've propagated a rng and seed keyword argument to all functions using rand, randperm or other random methods, and replaced all getRNG(seed) calls with rng = rng_from_rng_or_seed(rng, seed).

This took a lot of time / effort to fix all the tests, but upstream GraphRecipes has now reproducible tests.

@simonschoelly, this PR is now ready for review / comments.

[t-bltg]

I think I have addressed all the comments. Upstream tests ✔.

[simonschoelly]

Great, thanks a lot!

[t-bltg]

Amazing, thanks for your valuable comments (and obviously for your time !) 🎉 .