perf: track active requirement parents

[baszalmstra]

Instead of iterating over all requires_clauses in decide(), maintain an IndexSet of variables that are assigned true and have requires clauses. This reduces the iteration scope to only relevant variables. The idea is that this can make a big difference.

[baszalmstra]

I benchmarked this on my Windows Am64 machine. The results are a little underwelming.

There doesnt seem to be a huge difference, either positive or negative. But when I looked closer to the statistics I found that some of the worst case did actually improve significantly.

Summary: 577 faster (green), 423 slower (red)
Outliers: 6 packages improved by more than 10s, 0 got worse by more than 10s

Top 5 biggest improvements:
  - r-pgirmess, bson, jinja2-simple-tags, xa | -56.64s faster
  - cykhash, mujoco-simulate, kealib, mujs,  | -50.30s faster
  - armadillo, junit5, pytango-db, dagster-p | -32.96s faster
  - blist, yosys, django-autoslug, openorb-d | -14.22s faster
  - arabic_reshaper, cufflinks-py, opensm-co | -12.06s faster

56 seconds is huge! And this is actually also visible in the first graph. The number of cases that time out went from 0.5% to 0.2%!

This graph also shows the difference better (everything is capped to +10 or -10s improvement):

As you can see that some fast cases got faster and some got slower. But there are also a number of outliers that significantly improved.

Therefore I think this is a net improvement.