afterCompletion may be called before transaction _completes_ if reaper fires for a transaction and XAResource::end throws RuntimeException

EDITED BY TOM (to remove the report about the transaction not being disassociated):
If the reaper calls cancel on a transaction that is wedged, the transaction stays in 'aborting' state but it notifes Synchronizations that the tx has ended contrary to the spec.
This can be reproduced as follows:

• A TX is started and both a DB (last)resource and horntq (XA) resource get enlisted. The tx takes a long time and times out.
• Arjuna reaper thread notices the time out and starts a worker thread to cancel the TX.
• Before the worker thread can 'abort' the hornetq XA resource, the arjuna worker thread is interruped (txReaperCancelWaitPeriod expires) by arjuna
• Since the worker thread is interruped, the hornetq XA resource throws an 'HornetQInterruptedException'
• This unexpected exception causes arjuna to notify registered javax/transaction/Synchronization's (and return DB connection to the pool), without ending the Tx.

ORIGINAL DESCRIPTION:
We have noticed that arjuna leaks transactions when something 'unexpected' happens during tx aborting. The transaction stays in 'aborting' state while it does notify TxLIsteners that the tx has ended.

This can be reproduced as follows:

• A TX is started and both a DB (last)resource and horntq (XA) resource get enlisted. The tx takes a long time and times out.
• Arjuna reaper thread notices the time out and starts a worker thread to cancel the TX.
• Before the worker thread can 'abort' the hornetq XA resource, the arjuna worker thread is interruped (txReaperCancelWaitPeriod expires) by arjuna
• Since the worker thread is interruped, the hornetq XA resource throws an 'HornetQInterruptedException'
• This unexpected exception causes arjuna to notify registered javax/transaction/Synchronization's (and return DB connection to the pool), without ending the Tx.

From then on, jboss logs are full of:

Trying to start a new transaction when old is not complete: Old: < formatId=131077, gtrid_length=29, bqual_length=36, tx_uid=0:ffff40bab98c:1f71a485:5335ece7:3625a31, node_name=1, branch_uid=0:ffff40bab98c:1f71a485:5335ece7:3625bbc, subordinatenodename=null, eis_name=java:/XAOracleDS >, New < formatId=131077, gtrid_length=29, bqual_length=36, tx_uid=0:ffff40bab98c:1f71a485:5335ece7:3641efc, node_name=1, branch_uid=0:ffff40bab98c:1f71a485:5335ece7:36420a1, subordinatenodename=null, eis_name=java:/XAOracleDS >, Flags 0

Although the root cause is a misbehaving hornetq resource, I think arjuna should be a bit more resilient.

More background in https://access.redhat.com/support/cases/01061583/