Split the logic from ComputeSurfaceArea into a helper static function
and introduce a new method which computes triangle areas and writes the
resulting tensor into attributes of the mesh.

... and t::geometry::TriangleMesh::GetNonManifoldEdges methods.
The methods are defined in geometry::TriangleMesh API.

t::geometry::TriangleMesh::GetNonManifoldEdges mimics the logic of the
legacy method.

t::geometry::TriangleMesh::RemoveNonManifoldEdges follows the logic of
the legacy method but there are a few differences:
* the main difference is that the outer while-loop is removed. I don't
  see how after the first iteration any edge can have more than 2
  adjacent triangles, which makes the further iterations unnecessary.
* I count triangles with non-negative areas immediately and do not rely
  on the total number of adjacent triangles (which would also include
  triangles marked for removal).
* To choose a triangle with the minimal area out of the existing
  adjacent triangles I use a heap structure.