Kudu scan bottleneck due to sharing a single Kudu client for multiple tablet scans

This issue was observed when impala queries large datasets resides in Kudu. Even single ImpalaD is scanning multiple kudu tablets, it shows a slowness to retrive data eventhough ImpalaD makes parrellel scans. Reason for this is ImpalaD only uses a single Kudu client for multiple scans but KuduScanner::NextBatch runs on a single thread. So it's rpc reactor thread utilizes upto a single core and bottlenecks all parrelel scans. 

This behaviour makes Impala clusters that scans kudu cannot be vertically scales to the maximum performance/cores of a node.

Please refer the screenshots from Kudu slack channel for more information.