adapter: let the synchronous cut-over change the replication factor

[aljoscha]

Stacked on #38112. Addresses both of @ggevay's review comments on #38103.

Let the synchronous cut-over change the replication factor

Changing the factor while a reconfiguration is in progress is refused, because
the record captures a target factor when written and the controller's cut-over
writes it back to the realized config a tick later. A factor change applied
independently in the meantime gets silently clobbered.

That reasoning doesn't reach the synchronous cut-over: it folds the factor into
the target it transacts and retires the record in the same transaction, and no
consumer reads a retired record's target (desired_replicas, target_hydrated
and the controller's own cut-over all sit behind is_in_progress()). There's no
later cut-over left to do the clobbering, so it's exempt.

The refusal cost the escape hatch its most useful shape: forcing a wedged resize
through while scaling down had to be two statements, and in between the cluster
ran the old factor at the new size — the most expensive combination available,
exactly while someone is firefighting. fold_reconfiguration_target already
computed the right answer for a re-targeted factor; only the guard made that
branch unreachable.

The exemption is narrow. A zero-timeout rollback still leaves a record in
flight to settle on a tick, so it keeps the refusal. The error hint now offers
the cut-over as a third way out, which is useful precisely because after this
change the hint only ever fires when you didn't ask for one.

Test coverage was the reason this held: the existing size-and-factor case runs
with no record in flight (its own comment says so). The cut-over section now also
wedges a record, asserts the refusal still fires without a WAIT, then forces a
combined size-and-factor cut-over and checks both dimensions land, the replica set
converges, and the record settles cancelled — the record's own target factor was
abandoned, so cancelled rather than finalized is correct.

Drop Op::UpdateClusterReplicaConfig

Deleting the staged reconfiguration machine removed its last producer. The
variant, its transact.rs apply arm and its ddl.rs audit match arm were dead.
Confirmed no producers remain in src/ or test/.

Not in scope

@ggevay's comment prompted a wider question: should a factor change be allowed
while a reconfiguration is in progress generally, not just on the cut-over? The
factor is already a re-targetable dimension at record creation, and
mz_cluster_reconfigurations.changes plus the SHOW CLUSTERS summary already
render replication factor to N — so the fold, the durable target and the
user-facing surface all support it, and only re-targeting an existing record is
refused.

That's a user-visible semantics change on a surface about to go GA, with at least
three coherent designs that differ in what the user pays, so it's tracked in
SQL-626 rather than smuggled in here. This PR is a bug fix with no semantic
change: it's correct under every one of those designs.

[aljoscha]

Superseded by the controller-only design in #38103. The synchronous sequencer cut-over this PR extended has been removed entirely, so its replication-factor exemption is no longer valid. The useful dead Op::UpdateClusterReplicaConfig cleanup was folded into #38103. GitHub marked this PR merged automatically when its head became identical to its base during the stack rewrite. No change from this PR merged independently.