Improve Latency of Short Queries with Parallel Replicas by Offloading Index Analysis to Coordinator

During query execution with parallel replicas and enabled local plan, do not perform index analysis on workers. The coordinator will choose ranges to read for workers based on index analysis on its side (on the query initiator). This makes short queries with parallel replicas have as low latency as single-node queries. #72109 (Igor Nikonov).