Remeshing logic that "fixes" the results of corefinement operations to avoid an explosion of number of faces that hangs many models.

fast-csg-remesh operates as follows (needs fast-csg):

• A corefinement visitor keeps track of faces being split and copied (as suggested by @sloriot in Add coplanar decimation CGAL/cgal#5461 (comment)), giving each face a unique patch id that's inherited over copies and splits.
• If no face was split we skip the remeshing (note: we could force remeshing of all meshes, maybe when importing them).
• Maximal patches of neighbouring coplanar faces are then found with a flood-fill algorithm starting from descendants of split faces + exact coplanarity tests.
• Patches that contain holes are skipped for now
• Patch borders are found and we mark their collapsible vertices (those between collinear edges). If exactly two neighbour patches agree, those vertices will be dropped
• Each patch is replaced with a polygon that espouses its borders, and we let CGAL triangulate it.
• A new mesh is created with the edits, as in-place edits seem tricky.

Note: supersedes #4107 as by design it should lead to less coplanarity tests (for many faces we know they're coplanar because we know they have a common ancestor thanks to the corefinement visitor). Also this means we're not tied to the upcoming release of CGAL.

Some examples:

And some simpler ones:

• This goes from 644 faces to 12:

for (i=[0:2],j=[0:2],k=[0:2])
  translate([i*0.8, j*0.8, k*0.8])
    cube(1);

• This goes from 40 faces to 28

cube(1); translate([0.5, 0.5, 0]) cube(1);

• This goes from 44 to 12 (instead of 12 w/ nef, which is able to figure out it's now just a "cube"):

  cube(1); translate([0.5, 0, 0]) cube(1);

Known limitations

• No support for remeshing of faces with holes (as there are in the models in fast-csg performance regressions when combined with trust-corefinement and/or lazy-union #4114). Will probably need an exact version of GeometryUtils::tessellatePolygonWithHoles
• Currently only remeshing exactly coplanar faces (the example below with separately-rotated cubes has infinitesimal rounding errors due to transforms not using exact numbers).
• Still some bugs, e.g. trying to remove too many vertices in For Dyson Cinetic (@MichaelAtOz)
• The corefinement visitor slows down all operations a bit and uses a bit more memory (although some care was taken to allocate data structures sparingly) but that's quickly recouped in complex models
• This PR depends on Experimental sort-stl feature for stable outputs (fixes #420) #4146 as it adds STL export tests, this is (the main reason) why the tests aren't green yet.
• The resulting triangulation is sensitive to the order of the corefinement output, which is currently altered by optimization flags (Different corefinement output in debug vs. release CGAL/cgal#6363). Because of this, a fast-csg-remesh-predictibly feature was introduced that takes quite a bit longer to run (+30% total run time on the minkowski example mentioned above) but always starts each face from the "smallest" vertex. It is used by tests in conjunction with the sort-stl feature I've recently introduced (Experimental sort-stl feature for stable outputs (fixes #420) #4146) so we can test stable remeshing outputs against golden STL files.