This large commit adds a framework to generalize OSM data using
different strategies and implements several strategies. The source is
all in the new `src/gen` directory. It includes a new command
`osm2pgsql-gen` to access this functionality. Future versions might
integrate it with the `osm2pgsql` command, but because this is all
experimental it is kept separate for now. Call `osm2pgsql-gen` with
`-h` to get usage help.

See also https://osm2pgsql.org/generalization/ and the chapter on
Generalization in the manual:
https://osm2pgsql.org/doc/manual.html#generalization

For the raster support this adds two new library dependency: CImg and
potrace. Both have been around for a long time and are readily available
on all systems.

The following strategies work on a tile-by-tile basis and operate on
polygons:

The "vector-union" strategy buffers and unionizes polygons using vector
operations.

The "raster-union" strategy does a similar thing but does it in raster
space which is much faster. First the polygons are rendered into a
raster, an open/close operation is called (which basically does the same
thing as the buffering in vector space) and finally the resulting raster
is vectorized again.

The "builtup" strategy is intended to derive a layer of builtup areas
from landuse=residential/industrial etc. as well as building cover and
dense road networks.

The following strategies always work on all data (not tile-based):

The "discrete-isolation" strategy rates places based on some importance
metric to get a more even distribution of places on the map.

The new "rivers" strategy finds important rivers, this is still very
much work in progress.