With this change, users can submit a Flink job to a YARN cluster without having a local client monitoring the Application Master or job. You can basically fire and forget a Flink job to YARN.
For supporting this, the ApplicationMaster can now monitor the status of a job and shutdown itself once it is in a terminal state.

The change also verifies that various ways of setting the parallelism on YARN are passed through the system correctly (per job, session).

There was a bug in YARN container creation which made the configuration values for the heap offset useless. This change fixes this error.

All mentioned features and bugs are covered by the flink-yarn-tests.

This closes #542