Streaming window functions and optimization of RRF queries.

Window functions in MariaDB are currently always materialized into temporary tables before producing any results, even for simple cases, like ranking with a limit. This prevents index usage, enforces unnecessary sorting, and introduces avoidable I/O overhead. For hybrid search queries that combine vector and full-text search using Reciprocal Rank Fusion (RRF), this limitation forces both result sets to be fully computed before any top results can be returned, significantly impacting performance. This project introduces a streaming execution path for window functions, allowing eligible queries to produce rows incrementally without materialization. Building on this, a dedicated execution node for RRF will merge vector and full-text search streams and return top results early, avoiding full computation of either source and significantly improving hybrid search efficiency.