fix the issue that the client got wrong jobTrackerUrl when resourcemanager has ha instances

When the clusters set HA ResourceManagers , the conf  `yarn.resourcemanager.address.rm1` or `yarn.resourcemanager.address.rm2`  will replace  the conf `yarn.resourcemanager.address`. 

the conf `yarn.resourcemanager.address` may be not set,  then the method Hadoop23Shims.getJobLauncherRpcAddress  will return a wrong value. 

Maybe it should return the value of  the conf `yarn.resourcemanager.address.rm1` or `yarn.resourcemanager.address.rm2`