Regression tests for SingleMemoryStorageSchedule checkpoint clearing

[sghelichkhani]

Companion to dolfin-adjoint/pyadjoint#248, which fixes the SingleMemoryStorageSchedule checkpoint clearing reported in dolfin-adjoint/pyadjoint#211 and dolfin-adjoint/pyadjoint#260. Two regression tests, one per failure mode.

The first covers #211: a block variable reused only every third timestep is absent from the immediately preceding step's adjoint_dependencies, so the old condition cleared its checkpoint during forward replay and the reverse pass produced a wrong gradient (73765 instead of 3205 in the test). The test tapes ten such staggered timesteps and compares functional and gradient under SingleMemoryStorageSchedule against an unscheduled tape.

The second covers #260: a variable that is never redefined keeps the live block variable of its Function, and the clearing discarded its checkpoint at its last forward use even though the adjoint of that step still needs it as a linearisation point. The saved_output fallback to the live Function then supplies the taping-time value, so the gradient is wrong whenever the functional is re-evaluated at a new control first (140.5536 instead of 187.4048 in the test). The test runs two evaluate/derivative cycles at new controls, the second of which exercises the clearing that resumes once the adjoint dependencies have been revised by the first reverse pass.

In both tests the ordering is load-bearing and commented: the clearing only runs during the forward replay triggered by re-evaluating the reduced functional, so Jhat is called before derivative(). While test_burgers_newton parametrises over "SingleMemory", it never enables that schedule, so no SingleMemory validity comparison existed in the suite.

The TEMPORARY commit points pyadjoint-ad at the #248 branch so CI exercises the fix; it should be dropped, restoring the version pin, once #248 is merged and released. Since that branch sits on top of dolfin-adjoint/pyadjoint#257, CI will also show the pre-existing clear-down assertion failures that #5157 addresses, which are unrelated to these tests. As with the #257 companion, this should presumably end up on release once the ordering is settled.