Fix webapp worker thread and classloader leak on undeploy

[fangornoftheforest]

Coauthored by Claude

Three issues prevented clean shutdown of Argon2Function and BalloonHashingFunction executors in a servlet container lifecycle:

1. createExecutorService() registered a new Runtime shutdown hook on every call. Webapps that redeploy accumulated these hooks indefinitely; each hook pinned the undeployed webapp's classloader via its lambda, preventing garbage collection.

2. No public API existed to stop the workers. Runtime shutdown hooks only fire on JVM exit, not servlet contextDestroyed, so Tomcat's clearReferencesThreads consistently warned about password4j-worker threads on undeploy.

3. Even if a caller shut the executor down via reflection, the Argon2Function/BalloonHashingFunction INSTANCES caches still held references to instances whose executors had terminated, breaking subsequent hash calls.

Changes:

• Utils: single managed shutdown hook plus a static registry of executors. Registration is idempotent (guarded by a volatile hook reference).
• Password.shutdown(): public static entry point to be called from ServletContextListener.contextDestroyed. Shuts down every tracked executor, clears the function-instance caches so future calls rebuild cleanly, and removes the JVM shutdown hook to free the classloader.
• Argon2Function/BalloonHashingFunction: package-private clearInstances() used by the shutdown path.
• PasswordTest: regression test covering thread termination and reuse after shutdown.

Fixes #171: Failed to stop password4j-worker thread