Remove the unfixable testStateMachineFull and fix the dead replay pool

[vharseko]

Context

StateMachineTest.testStateMachineFull — a walk through all replication state machine transitions (NC→D→N→NC→N→D→NC→D→N→BG→NC→N→D→BG→FU→...) — has been disabled since the ForgeRock era. A deep attempt to revive it (part of the slow group audit) showed it cannot work in its historical form:

• The degraded status only appears while the RS queues changes faster than the DS replays them (checkDSDegradedStatus samples the RS-side queue every 5 seconds against a threshold of 1). On modern hardware the publisher and the replay run at comparable rates, so every degraded transition is a race: bounded change batches get replayed before the analyzer samples a non-empty queue, and unbounded floods exhaust the 512m test JVM heap with real backend entries first (reproduced consistently).
• On 2009-era hardware the replay was durably slower than the publisher, which is what the scenario silently relied upon.

The test and its private helpers (resetGenId, BrokerInitializer) are removed with an explanatory note in the class; the degraded/normal transitions remain covered deterministically by testStateMachineStatusAnalyzer.

Real bug found and fixed along the way

The class helper invoked applyConfigurationChange() with a bare Mockito mock. mock(ReplicationSynchronizationProviderCfg.class).getNumUpdateReplayThreads() returns 0 (Mockito's default for Integer, not null), so the null-check in getNumberOfReplayThreadsOrDefault() never fired and the call silently killed the shared replay thread pool of MultimasterReplication for every test in the class — a regression from the refactoring that replaced MultimasterReplicationFakeConf with a bare mock (thread dumps showed zero Replica replay threads; the DS listener queued updates nobody replayed, window acknowledgements never went out and the RS writer froze on acquirePermitInSendWindow).

• The mock is now stubbed to use the default pool.
• Product hardening: getNumberOfReplayThreadsOrDefault() falls back to the default for non-positive values — the configuration schema enforces a minimum of 1, so a zero can only come from a broken caller and would silently disable replication replay.

Verification

StateMachineTest: 2 consecutive runs, 5/5 green (~125s) — testStateMachineStatusAnalyzer now exercises a live replay pool for the first time.

[maximthomas]

I think this line should be removed

Suggested change

/*