fix(engine): prevent record corruption during message TTL checking

When multiple message TTL checker were running concurrently, for example
because one broker is leader for multiple partitions, the scheduled
jobs tried to write the same message record concurrently. Writing
records (or really, `UnpackedObject`s) is not thread safe. Using a
static instance of the empty message record would result in
deserialization errors while trying to create a copy before writing to
the logstream.
Usually this would kill the processing actor and stop processing for one
partition. With `enableMessageTtlCheckerAsync` enabled, it would only
prevent message TTL checking until a new instance of `MessageObserver`
is started.

(cherry picked from commit f318044)

engine/src/main/java/io/camunda/zeebe/engine/processing/message/MessageTimeToLiveChecker.java

@@ -33,7 +33,7 @@
  */
 public final class MessageTimeToLiveChecker implements Task {
 
-  private static final MessageRecord EMPTY_DELETE_MESSAGE_COMMAND =
+  private final MessageRecord emptyDeleteMessageCommand =
       new MessageRecord().setName("").setCorrelationKey("").setTimeToLive(-1L);
 
   /** This determines the duration that the TTL checker is idle after it completes an execution. */
@@ -88,7 +88,7 @@ public TaskResult execute(final TaskResultBuilder taskResultBuilder) {
 
               final boolean stillFitsInResult =
                   taskResultBuilder.appendCommandRecord(
-                      expiredMessageKey, MessageIntent.EXPIRE, EMPTY_DELETE_MESSAGE_COMMAND);
+                      expiredMessageKey, MessageIntent.EXPIRE, emptyDeleteMessageCommand);
               return stillFitsInResult && counter.incrementAndGet() < batchLimit;
             });