Moving discussion from #983 (comment)

In current Shapely, there are two attributes that both return the GEOS pointer:

In [7]: from shapely.geometry import Point

In [8]: p = Point(1, 1)

In [9]: p
Out[9]: <shapely.geometry.Point POINT (1 1) >

In [10]: p.__geom__
Out[10]: 94617737200752

In [11]: p._geom
Out[11]: 94617737200752

In the branch subclassing from the C extension type (#983), there is now a third attribute inherited from PyGEOS that does the same:

In [12]: p._ptr
Out[12]: 94617737200752

I think it is fine to keep some way to get access to the GEOS pointer (eg if you want to use numba or cython yourself in some application where you know you are using the same GEOS version as Shapely, and thus passing along pointers can be safe). So we should maybe keep one of those attributes (although for cython, a function to get the GEOS pointer from a python object could also be used for this instead of an attribute).
But we don't necessarily need to keep all of them, and we can deprecate others?

As @mwtoews mentions at #983 (comment), __geom__ is not an official "dunder" method of python (and here also not used in a context for which typically dunder methods are defined), we could maybe deprecate __geom__ in Shapely 1.8, keeping _geom in Shapely 2.0 (since that is backwards compatible with older Shapely versions)?

cc @sgillies @caspervdw