Fix EntityManager leak in ThreadLocal after failed transaction

[SantanuKar43]

This PR fixes a bug in JpaLocalTxnInterceptor where a broken EntityManager could remain in a thread’s ThreadLocal after a failed transaction, leading to subsequent requests failing when the same thread is reused by the web server (e.g., Jetty).

Problem

• Guice Persist stores the EntityManager in a ThreadLocal.
• If a transaction fails due to a database connection issue (e.g., network blip):
• txn.isActive() incorrectly returns false because the connection is already dead.
• The cleanup logic is skipped due to this check.
• The broken EntityManager is never removed from the ThreadLocal.
• When the thread is reused for another request, the stale EntityManager is returned, causing immediate failures.

Fix

• Updated cleanup logic to always remove the EntityManager from the ThreadLocal if this interceptor started the work, regardless of txn.isActive().
• Reorganized the try/catch/finally blocks so that commit/rollback semantics remain correct while ensuring cleanup runs after transaction handling.
• Ensures no broken EntityManager instances leak into future requests.

Related Issue

Fixes #1916 and #1179

[SantanuKar43]

@sameb @lukesandberg can you please review this?