Performance degradation for massive deployments in Weld subsystem

This issue is a copy of EAPSUP-316 which points out deployment time increase for massive applications (EAR with 500 EJB jars in this case - yea, I know, your daily deployment sample ). There is no real drawback for any small deployment, the difference would be negligible.

Turns out a lot of time is spent inside Weld subsystem which needs to analyze the accessibility graph but also process external bean archives. While I don't see how to easily improve the former, the latter can be cut down significantly with a simple cache.

See the original issue for more details, I will also re-attach the reproducer here for ease of double checking.