Keep the Vector Set replay channel alive across AofProcessor disposal

[tiagonapoli]

Stacked on #1996 (which is stacked on #1995) — review those first; this PR targets tiagonapoli/vs-2-recovery-reentrancy, so its diff shows only this change.

Problem

AofProcessor.Dispose() called VectorManager.ShutdownReplayTasks(), which is a permanent shutdown — it completes the replication replay channel writer and parks the replay tasks for good.

But the VectorManager belongs to the database, not to the processor, and a node constructs and disposes several AofProcessors over its lifetime. SingleDatabaseManager.ReplayAOF builds a transient one during startup AOF recovery.

So on any node started with Recover enabled — a normal production replica config — the channel was already closed before the node began serving. From that point HandleVectorSetAddReplication could no longer queue:

var queued = replicationReplayChannel.Writer.TryWrite(...);
if (!queued) replicationBlockEvent.Decrement();   // no log, no error

It discards whatever it cannot queue, silently. The node dropped every VADD replicated to it while continuing to report healthy AOF offsets — silent data loss with no signal at all.

Fix

Dispose() now only drains. VectorManager.Dispose() still performs the real shutdown when the database goes away.

Tests

Adds ClusterVectorSetRecoveryTests, which restarts the replica, the primary, and both nodes with tryRecover: true — the configuration that exposed this.

Note these tests also depend on #1996: without it, an aborted recovery leaves the store corrupt enough that a subsequent read dies with AccessViolationException in TsavoriteKV.TraceBackForKeyMatch, taking down the test host. That is why this is the top of the stack.

Validation

• 19/19 pass with the fix (full suite, top of stack).
• All 3 recovery tests fail without it.
• Regressions green: Garnet.test.cluster.vectorsets 80/80, replication.disklesssync 28/28, replication 103/103. ShutdownReplayTasks() is left public but now has no callers.