YARN service containers to support different memory limit policies

Service containers may have different memory policies from applications. Specifically, it may be more important to keep the container running if its memory goes over allocated, rather than killing it.

If the memory over-use could be detected, the AM could then make an attempt to request more resources for that container, so increasing its allocation. And, if that fails, shutting down the node cleanly. As it is, a killed container looks like a failure to the AM, which runs the risk of concluding that the app and/or host is in trouble, when it is merely the memory requirement is too low