What are your thoughts about the preparation interface? We need to be able to take a reliable path for preparations which might have RTree but not MonotoneChain, for example...

I'm not sure really! I'm hoping that will be easier when there is a concrete PR using it. Probably we will need to make some changes to some of this too.

Sampling could use this - e.g. to_edges could access an EdgeDecomposition preparation, and linestrings could have precomputed per-segment arclength, multipolygons can have precomputed area per polygon, triangulation, etc.

A current list of preparations for #184 would be:

• EdgeVector: A vector of edges, per geometry.
• Triangulation: A tuple of (points, faces), per polygon. Multipolygons are at a higher level than this and don't care.
• SegmentLengths: A vector of length per edge, linestring only.
• SubpolygonArea: A vector of area per subpolygon, multipolygon only.

I guess each preparation would have a specialized function? Or we can override e.g. area on multipolygons which have the SubpolygonArea preparation.

Not sure how to dispatch on only prepared geoms which have some preparation, nor how to define an order of priority for that dispatch.