Template distribute_entity_data to support multiple data-types

[jorgensd]

First step towards resolving #1253
Second step would be to template XDMFFile.read_meshtags. This is fairly trivial, but I figured we should agree on the core changes before doing that change.

Currently tested with:

from mpi4py import MPI
import numpy as np
import dolfinx

mesh = dolfinx.mesh.create_unit_square(MPI.COMM_WORLD, 2, 2)
num_vertices = mesh.topology.index_map(0).size_local
entities = np.arange(num_vertices, dtype=np.int32)


def distribute_data(dtype):
    values = entities.astype(dtype).copy()
    if dtype == np.float64:
        values -= 0.6
    return dolfinx.io.distribute_entity_data(mesh, 0, entities.reshape(-1, 1).copy(), values[::-1].copy())


a = distribute_data(np.int32)

b = distribute_data(np.float64)
c = distribute_data(np.complex128)
print(a, b, c)

I've removed the unnecessary mesh geometry templating of this function, and made the Python-binding more low-level (improving the Python layer of distribute_entity_data to work with the new binding). The PR also has documentation improvements of this function.

This would enabling reading vertex, facet and cell data that is single or double precision, which is the case in many biomedical applications (where one for instance have voxel data, or material parameters).

The major difference in the code is that the communication of entity indices and values are split into two separate neighbourhood alltoallv calls.

No major changes in the actual functional structure of the code.

[jorgensd]

@massimiliano-leoni could you test this?:)

[garth-wells]

I would suggest leaving the code in the .cpp file and explicitly instantiating for the small number of types that we'll use (at this stage). It's preferable to keep code in the implementation file. This is what we do elsewhere for the mesh.

[jorgensd]

@garth-wells any further changes you would like to see in this PR?