Make sure containers are not allocated to 'incorrect tasks'

Currently, containers are re-used at the same priority level - irrespective of JVM options, local-resources etc - which implies different vertices with the same priority can end up re-using containers.
If there's multiple initial map vertices - they can end up re-using containers and reading incorrect split information.
This will be a stopgap fix - until TEZ-324 is fixed. Likely something as simple as blocking re-use across vertices.