[Feature] Implement ACVD

[csparker247]

Description

Implement Approximated Centroidal Voronoi Diagrams (or similar). Not a flattening technique, but useful for mesh conditioning and coarsening, important preprocessing steps for parameterization.

Caveats

There is a reference implementation of this library available online. We want this to be a pure implementation from the reference papers, so we MUST NOT examine this reference library.

References

Valette, Sébastien, and Jean‐Marc Chassery. "Approximated centroidal voronoi diagrams for uniform polygonal mesh coarsening." Computer graphics forum. Vol. 23. No. 3. Oxford, UK and Boston, USA: Blackwell Publishing, Inc, 2004. [PDF]

Sébastien Valette, Jean-Marc Chassery, Rémy Prost. Generic remeshing of 3D triangular meshes with metric-
dependent discrete Voronoi Diagrams. IEEE Transactions on Visualization and Computer Graphics, 2008, 14
(2), pp.369–381. ⟨10.1109/TVCG.2007.70430⟩. ⟨hal-00537025⟩ [PDF]

[csparker247]

Scoping note: the ACVD engine will live in libcore, tracked in educelab/libcore#22. This issue
becomes the thin adapter/wrapper on top of it.

Rationale for the split:

• ACVD needs no sparse linear algebra — dense 3×3 is the largest system it ever solves (the one
  exception is a self-contained symmetric eigensolve for the anisotropic 2008 layer). libcore can
  host all three papers without gaining a dependency, whereas hosting the engine here would make
  Eigen the price of admission for any consumer that only wants remeshing.
• ACVD never mutates topology: it reads adjacency, writes one integer label per item, and emits a
  new mesh. So HalfEdgeMesh isn't needed for it — and its shared_ptr elements, multimap edge
  store, outgoing_edges() returning a vector<EdgePtr> by value, and wheel() throwing on
  boundary vertices all argue against it for the inner loop of a large minimizer. The engine works
  on flat indexed arrays either way.
• Dependency direction stays OpenABF → libcore (heavy → light).

The wrapper here is small: HalfEdgeMesh vertices already carry idx, so building the adjacency
is one pass over faces(), and the result comes back through insert_vertices/insert_faces.
Worth noting this repo also has the benchmark meshes (up to 400k faces) that the engine's
performance work will want.

The clean-room constraint in the Caveats above carries over to libcore#22 and is restated there.