Use quadtree for .contains_properly

[thomcom]

Description

Closes #889
Closes #896

This PR makes a number of changes to the binpred system, including adding boundary exclusion for .contains_properly to the quadtree line-crossings algorithm.

I've closed and then reopened this PR due to benchmark results that showed that using pairwise and column-limited forms of point-in-polygon don't offer any performance improvement but would just require more complicated code paths.

Checklist

• I am familiar with the Contributing Guidelines.
• New or existing tests cover these changes.
• The documentation is up to date with these changes.

[harrism]

One option would just be to change this output format before you consume it here. For the small-polygon-count version, just convert the bitmask into the same format output as the quadtree (all pairs): one row per intersection, containing a polygon index and a point index.

Then contains_properly only needs to convert the results if the user requests pairwise instead of all-pairs.

[thomcom]

Turns out it is really easy to change the output format with a cudf.stack() call, so I'm investigating this approach to reduce code-paths.

[thomcom]

I've finished that refactor, it makes solving the discrete math problem at the end harder but I'm ok with it for the current iteration of the architecture.

[thomcom]

Please approve, this is ready for merge. I realized that the fundamental reason I've been grinding my teeth on this at all is because I keep trying to add small exceptions or refactorings to an architecture that isn't designed for it yet. I need to merge geom_equals and contains_properly before I sit down for a few days and redesign the architecture based on the lessons that have been learned building the first two basic predicates.

[thomcom]

/merge