What architectural decisions enabled pg_textsearch to outperform Tantivy/ParadeDB by 4.7×, especially given Tantivy’s reputation for speed?