[ycb] Update inertias to use constant-density estimations from meshes

[jwnimmer-tri]

Transcribed from RobotLocomotion/drake#10704:

---

Follow-up from: RobotLocomotion/drake#10702

We should update the models to use an estimation from the mesh (if possible).

This may involve:

• Simplifying / separating the inertial estimation from the current //manipulation/util:meshlab_to_sdf binary (e.g. just give it an OBJ, and it'll give you an <inertial/> with no frame adjustment)
• Ensuring that the center of mass is accurate given the inertial adjustments.

[nepfaff]

What about something like the following: https://github.com/RussTedrake/manipulation/blob/master/manipulation/create_sdf_from_mesh.py#L20

or

mesh_o3d = o3d.io.read_triangle_mesh(args.mesh_path)
    vertices = np.asarray(mesh_o3d.vertices)
    triangles = np.asarray(mesh_o3d.triangles)

    triangles_drake = np.array([SurfaceTriangle(t[0], t[1], t[2]) for t in triangles])
    mesh_drake = TriangleSurfaceMesh(triangles_drake, vertices)

  # TODO: Compute density from mass and volume
    inertia = CalcSpatialInertia(mesh=mesh_drake, density=density)

I found these to work quite well in the past, even when the meshes aren't entirely (but close to) watertight.

[jwnimmer-tri]

We would probably just run https://drake.mit.edu/pydrake/pydrake.multibody.mesh_to_model.html and manually tweak from there.