BUG: Incorrect local to global map for non-Lagrange quadrilaterals

[pbrubeck]

Describe the bug
It appears that we are failing to get the facet orientations correctly on non-Lagrange bases for Q. But this only happens in parallel if the mesh distribution is not overlapping.

Steps to Reproduce
Steps to reproduce the behavior:

Here's an MFE assembling a Form on a quadrilateral mesh with two elements using Lagrange (variant="spectral") and non-Lagrange (variant="integral") with the dfault overlap (FACET) and without overlap (NONE).

from firedrake import *
from firedrake.petsc import PETSc

degree = 3
overlaps = {"NONE": (DistributedMeshOverlapType.NONE, 0),
            "FACET": (DistributedMeshOverlapType.FACET, 1),}
variants = ("spectral", "integral")

for variant in variants:
    for overlap in overlaps:
        distribution = {"overlap_type": overlaps[overlap]}
        mesh = UnitSquareMesh(2, 1, quadrilateral=True, distribution_parameters=distribution)
        V = FunctionSpace(mesh, "Lagrange", degree, variant=variant)

        c = assemble(TestFunction(V) * dx)

        # Zero out interior dofs
        own = mesh.cell_set.size
        interior = V.finat_element.entity_dofs()[2][0]
        c.dat.data[V.cell_node_list[:own, interior]] = 0

        # Zero out boundary dofs
        bc = DirichletBC(V, 0, "on_boundary")
        bc.zero(c)

        c.dat.data[abs(c.dat.data) < 1E-10] = 0
        with c.dat.vec as vec:
            PETSc.Sys.Print(variant, distribution, vec.norm())
            vec.view()

Output of mpiexec -n 2 python script.py:

spectral {'overlap_type': (<DistributedMeshOverlapType.NONE: 1>, 0)} 0.04910463758239919
spectral {'overlap_type': (<DistributedMeshOverlapType.FACET: 2>, 1)} 0.04910463758239919
integral {'overlap_type': (<DistributedMeshOverlapType.NONE: 1>, 0)} 0.12500000000000014
integral {'overlap_type': (<DistributedMeshOverlapType.FACET: 2>, 1)} 0.1767766952966371

Expected behavior
The norm of the assembled Cofunction should be independent of the mesh distribution, but it should still depend on the variant.

Error message
No error message.

Additional Info
The MFE zeros out all entries except for the interface dofs on the interior of the domain. For variant="integral", the overlapping case gives the correct solution [0., -0.176777], whereas it appears that the non-overlapping case gives [-0.088388 -0.0883883]. This suggests that we are failing to map the contributions coming from each cell with the right entity permutation.