Lazy load networkx in topology modules

[mattwthompson]

I was profiling some other code and found that a ~100 ms import of networkx happens in the toolkit whether or not it's actually used. I remember lazy-loading this in the past, but that's reflected in the current code.

This way of profiling here shows roughly a 50-100 ms speedup depending on how one squints their eyes

$ git checkout upstream/main
$ hyperfine 'python -c "from openff.toolkit import Molecule"' --prepare 'python -c "from openff.toolkit import Molecule"'
Benchmark 1: python -c "from openff.toolkit import Molecule"
  Time (mean ± σ):      1.119 s ±  0.047 s    [User: 0.909 s, System: 0.142 s]
  Range (min … max):    1.085 s …  1.241 s    10 runs

  Warning: Statistical outliers were detected. Consider re-running this benchmark on a quiet system without any interferences from other programs. It might help to use the '--warmup' or '--prepare' options.

$ git checkout lazy-load-networkx
Previous HEAD position was ef2f514f Drop py311 from conda tests (#1873)
Switched to branch 'lazy-load-networkx'
$ hyperfine 'python -c "from openff.toolkit import Molecule"' --prepare 'python -c "from openff.toolkit import Molecule"'
Benchmark 1: python -c "from openff.toolkit import Molecule"
  Time (mean ± σ):      1.050 s ±  0.048 s    [User: 0.864 s, System: 0.121 s]
  Range (min … max):    1.007 s …  1.143 s    10 runs

With this patch, here's a rough picture of what takes so long to run this import. There are some other obvious candidates for lazy-loading, but I'm not going for them now:

• Tag issue being addressed
• Add tests
• Update docstrings/documentation, if applicable
• Lint codebase
• Update changelog

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 93.70%. Comparing base (ef2f514) to head (5e37a6e).
Report is 48 commits behind head on main.

Additional details and impacted files

[j-wags]

Thanks @mattwthompson - Please update the releasenotes before you merge :-)