Extend yarn.nodemanager.delete.debug-delay-sec to support application level.

Now the yarn.nodemanager.delete.debug-delay-sec is a cluster level setting.

In our busy production cluster, we set it to 0 default for preventing local log boom.

But when we need deep into some spark/MR etc jobs errors such as core dump, i advice to support enable a job level setting for delay of deletion for local logs for reproduce the error.

wangda tangzhankun  xgong epayne

If you any advice about this support?