Faster transitive implementation

[camsaul]

Hey @weavejester,

I've noticed that lein-ring starts really slow for larger projects (for example it usually takes 15-20 minutes to launch Metabase via lein ring server, compared to ~1 with lein run).

I did some profiling and determined 98% of the launch time was spent in ns-tracker.dependency/transitive, mainly because it recursively calls transitive on everything before finally removing duplicates in seq-union. I changed this to an implementation that avoids recursive calls for dependencies that have already been dealt with, which has pretty dramatic performance improvements.

To test this I captured a real-world graph of ~200 namespaces from Metabase during launch and compared the old and new implementations. The namespaces graph and code I used to compare implementations is here if you'd like to see for yourself. When testing locally, the new implementation is around 1000x faster:

Implementation	Time
Old	873.714 ms
New	0.754 ms

I can also confirm that it has fixed the slow launch time issue I was running into.

The order of results in the new implementation are consistently breadth-first whereas results in the old implementation are mostly breadth-first but, not consistently so. I put an example comparison between the orders returned by the implementations here. As far as I can tell the slight order differences has no effect on the reloading code that relies on it.

[weavejester]

Thanks for the improvement! Can you change the commit message to:

Speed up transitive function

[camsaul]

@weavejester done