Don't use saved_output to create checkpoints for timesteps

[angus-g]

We have a few blocks, T (the BC function) <- AssignBlock (for T += 1), <- DirichletBCBlock <- AssignBlock <- DirichletBCBlock. At the end of timestep 0 (or the beginning of timestep 1), we have the first 3 of these blocks on tape. During the add_dependency phase of the solve, the DirichletBCBlock was also checkpointed, so its associated block variable has a _checkpoint member which is just the T function (https://github.com/firedrakeproject/firedrake/blob/a9f2c621633f0e56c69c9ab42c6f2343051f4839/firedrake/adjoint_utils/dirichletbc.py#L36-L43)

Now, when timestep 1 is due to begin, the process_taping method checkpoints all the blocks:

pyadjoint/pyadjoint/checkpointing.py

Lines 203 to 204 in 0b13480

	self.tape.timesteps[timestep - 1].checkpoint(
	_store_checkpointable_state, _store_adj_dependencies, self._global_deps)

This process goes through the saved_output property:

pyadjoint/pyadjoint/tape.py

Lines 839 to 840 in 0b13480

	for var in self.adjoint_dependencies.union(self.checkpointable_state):
	self._checkpoint[var] = var.saved_output._ad_create_checkpoint()

Let's consider when var here is the block variable associated with the DirichletBC. The saved_output property will use the existing checkpoint:

pyadjoint/pyadjoint/block_variable.py

Lines 57 to 62 in 0b13480

	@property
	def saved_output(self):
	if self.checkpoint is not None:
	return self.output._ad_restore_at_checkpoint(self.checkpoint)
	else:
	return self.output

This has the effect of calling DirichletBCMixin._ad_restore_at_checkpoint on a checkpointed function: https://github.com/firedrakeproject/firedrake/blob/a9f2c621633f0e56c69c9ab42c6f2343051f4839/firedrake/adjoint_utils/dirichletbc.py#L45-L48

Because of the use of set_value here, the DirichletBC is then modified (and refers to the previous value of T). It seems like the weak checkpointing in DirichletBC is probably correct, which points to the accessing of saved_output during TimeStep.checkpoint being incorrect. I'm really unclear about what's checkpointed and when and when to use .output vs. .saved_output, etc. I don't know if there are further implications for this change.