Low-level handling of degenerate polygons

[rhattersley]

Projection._project_polygon() is rather unhelpfully converting degenerate polygons (such as those produced by matplotlib.contourf) into global extent.
e.g.
Polygon([(20, 20), (20, 20), (20, 20), (20, 20)])
Polygon([(20, 20), (20, 20), (20, 20.1), (20, 20)])

Whilst it's easy to filter these out in the MPL integration layer, ideally the low level re-projection code should be able to deal with them.

[pelson]

Actually, it turns out that many of the most common clipping algorithms don't handle degenerate polygons (d3 suffers from this problem too: d3/d3#2025). Essentially, the cutting operation can be seen as an extension of a clip (you keep both sides instead of just one side), so picking an algorithm which handles degeneracies is critical. The truth is though, we could just as well pre-process degenerate "polygons" and either remove them entirely, or convert them to lines.