perf(query-engine): use incremental per-step merging for range queries #641

The shared range pipeline still creates a fresh WindowMerger and fully merges buckets for every output
timestamp instead of carrying merged state forward with WindowMerger::slide(). Measure and implement incremental per-step merging where semantics permit it, with correctness tests for Sliding, Tumbling, DeltaSetAggregator keys, missing buckets, boundaries, dual-
population keys, and top-k. DeltaSet-specific O(N^2) key replay remains tracked separately in #599. Related: #581 and #589.