Investigate faster polygon operations

[Huite]

The current burn_vector_geometry relies on calling earcut to break a polygon down into a triangular mesh, which is then checked for overlap. I don't necessarily see a better way for all_touched=True since it relies on computing overlaps. But in case of all_touched=False, it's just searching for centroids.

In this case, it's just a point in polygon check. The ad hoc benchmarking seemed to show that is was beating geopandas points-in-polygon, but not massively so (something like one order of magnitude).

I came across these two:

https://github.com/dengwirda/inpoly
https://github.com/KlausC/PolygonInbounds.jl

They do some prior work so it should be more efficient to search collections of points. It would be worthwhile to investigate, although it should maybe end up in numba-celltree instead.

[Huite]

The 10x speed up compared to geopandas is not that impressive -- I guess because it's building an intermediate celltree for the earcut polygons.

However, that tree could be re-used in case multiple operations are required for a polygon dataset. This a pretty convincing argument to add an earcut staticmethod which returns a UgridDataArray with the earcut topology as the mesh, and returns the polygon indices as the data.

This also gives a clean composing way of implementing: #234