Introduce create_identity_partitioner for refined mesh partitioning

[schnellerhase]

Introduces create_identity_partitioner which facilitates the exchange of ghost cells of a refined mesh whilst maintaining the partition of the coarse mesh.

This fixes (partially) #3443.

Additionally, this changes

1. move compute_destination_ranks from anonymous namespace to dolfinx::graph,
2. python dolfinx.mesh.refine 's partitioner default argument value was not aligning with cpp layer, fixed,
3. default argument for refine's option arg changed from Option::none to Option::parent_cell,
4. partitioner argument of refine is now std::optional, allows for handling of case of not provided (which is different from providing nullptr as callable).

[schnellerhase]

Verification code:

import febug
import pyvista
import dolfinx
import dolfinx.mesh
from mpi4py import MPI

mesh = dolfinx.mesh.create_unit_square(MPI.COMM_WORLD, 4, 4)
mesh.topology.create_entities(2 - 1)
mesh.topology.create_connectivity(2 - 1, 2)
mesh.topology.create_entities(1)
[refined_mesh, _, _] = dolfinx.mesh.refine(
    mesh, option=dolfinx.mesh.RefinementOption.parent_cell
)
subplotter = pyvista.Plotter(shape=(2, 1))

subplotter.subplot(0, 0)
febug.plot_mesh(mesh, plotter=subplotter)

subplotter.subplot(1, 0)
febug.plot_mesh(refined_mesh, plotter=subplotter)
# subplotter.show()
subplotter.save_graphic(f"out_{MPI.COMM_WORLD.rank}.pdf")

Output ($np=2$):
out_0.pdf
out_1.pdf

Output ($np=3$):
out_0.pdf
out_1.pdf
out_2.pdf

[schnellerhase]

Tests fail due to refinement routine now not being able to handle the case of no partitioner and no parent_cell flag being set anymore. This originates from the need to have parent_cell computed before the partitioner may be constructued. Should we disallow this combination?

Also, I have tested the implementation with the above code snippet - not sure how to extend that to a reasonable unit test though.

[schnellerhase]

I have a problem with the std::optional here in the Python layer. The export of the partitioner callback now has signature std::optional<CellPartitionFunction> (to allow for handling the different cases of nullptr and std::nullopt). Passing None results in a std::nullopt and no longer the nullptr as before . However I am not aware of another way to produce a nullptr in the C++ layer other than by passing None.

It comes down to: how can a std::optional<CellPartitionFunction> = nullptr case be produced form the python caller side?