BUG: domain of Constant not preserved in adjoint replay

[stephankramer]

Describe the bug
When assigning to a Constant T the value of a Constant S on a different domain, it appears that in the replay the checkpointed value of T after the assignment has taken the domain of S. I think this is because the checkpointed value of T just points to the checkpointed value of S (through a DeferredCheckpoint). This means T ends up on the wrong domain.

Steps to Reproduce
Consider the following example where S is on a normal FEM mesh, and T is on a vertexonly mesh mesh0. The functional integrates T over mesh0. This works correctly in the initial run of the model, but in the replay it errors in the assembly of that integral complaining NotImplementedError: Assembly with multiple meshes is not supported and digging into the 0-form it's trying to assemble at that point it does appear that the checkpointed value of T it is using there is on mesh instead of mesh0

from firedrake import *
from firedrake_adjoint import *
mesh = UnitSquareMesh(10,10)
mesh0 = VertexOnlyMesh(mesh, [[.5,.5]])
S = Constant(1.0, domain=mesh)
T = Constant(2.0, domain=mesh0)
T.assign(S)
J = assemble(T*dx(domain=mesh0))
rf = ReducedFunctional(J, Control(S))
rf(Constant(3.0, domain=mesh))

When constructing T immediately from the value of S:

T = Constant(S, domain=mesh0)

the same error is produced.