Fix quadratic-time ModuleNode.sort_types_by_inheritance

[swolchok]

At Meta, we have some large .pyx files generated by https://github.com/facebook/fbthrift . I found through profiling that ModuleNode.sort_types_by_inheritance was taking about 10% of the time for Cython to process our largest file, and through inspection that it took quadratic time. This PR changes sort_types_by_inheritance to usesorted to avoid quadratic time, with the result that sort_type_by_inheritance take a small fraction of a percent on the same large workload after this change.

This function seems to be covered by the cdef_class_order and ooo_base_classes tests; they failed during development and now pass.

[swolchok]

This needs further work; will update when it's ready.

[da-woods]

The failing tests are probably mostly an unrelated issue to do with a new version of line_profiler. So they may not be to do with this PR. I'll rerun the tests on this PR when that's fixed.

[swolchok]

The failing tests are probably mostly an unrelated issue to do with a new version of line_profiler. So they may not be to do with this PR. I'll rerun the tests on this PR when that's fixed.

thanks! I'm seeing failures on the internal version of this as well, though; I'll have to debug those and add more tests to Cython.

[swolchok]

The problem is the sort comparator I defined isn't a total ordering. I'll have to implement a proper topological sort.

[da-woods]

Closing and reopening to rerun CI

[scoder]

I simplified the implementation. @swolchok, could you please validate that this doesn't lead to a performance regression on your side again?

[scoder]

I'll merge this since it's a major improvement. We can have further PRs if we need more.

Thanks!