Cache the Kudu client between queries

Right now each time we analyze a new kudu query and create a new plan we create a new client instance, meaning it will query the master for tablet locations again. If we cache the client between queries it will keep that metadata internally (and invalidate if/when needed).