fix: skip writes and still shut down after failed DB manager startup

[arimu1]

Fixes #4241 and #4242

Description

AbstractDatabaseManager checks isRunning() on the shutdown path but not on the write path.

After startupInternal() throws, startup() logs the cause and leaves running = false. Then:

• write() did not check it — kept accepting events and dereferencing state startup never assigned (NPE per event, burying the original cause).
• shutdown() only ran shutdownInternal() when running — so resources acquired during a failed startup (or in the manager factory) were never released. Returns true anyway.

This change:

1. Makes write() return early when !isRunning(), logging a single status warning instead of one failure per event.
2. Makes shutdown() always invoke shutdownInternal() once (tracked independently of running), so partial startup state can be cleaned up.
3. Makes CassandraManager.shutdownInternal() null-safe when session was never created (motivating failure mode from the issue / related log4j-cassandra: failed appender startup leaks the DataStax Cluster, so the JVM never exits #4242).

Checklist

• Base your changes on 2.x branch if you are targeting Log4j 2; use main otherwise
• ./mvnw verify succeeds (the build instructions)
• Non-trivial changes contain an entry file in the src/changelog/.2.x.x directory
• Tests are provided

Testing

• JDK 17 (branch enforcer requires [17,18)) — JBR 17.0.14
• ./mvnw test -pl :log4j-core-test -am -Dtest=AbstractDatabaseManagerTest -Dsurefire.failIfNoSpecifiedTests=false — 11/11 pass
  • New: testFailedStartupSkipsWrite, testFailedStartupStillShutsDown, testShutdownWithoutStartupStillRunsShutdownInternal
• Spotless check clean on :log4j-core, :log4j-core-test, :log4j-cassandra

[vy]

@arimu1, could you take care of the build failures, please?

[arimu1]

Thanks for the review, @vy — CI should be green now.

AbstractDatabaseAppenderTest.testAppend was failing because the Mockito mock’s isRunning() is final and defaults to false, so write() returned early after the production guard. The test now uses a spy of a stub manager and calls startup() so the happy path still exercises writeThrough().

Failed-startup behavior remains covered by AbstractDatabaseManagerTest (testFailedStartupSkipsWrite, testFailedStartupStillShutsDown, etc.).