This part is easy - Silva 138.1 contains outdated taxonomy names that were updated in 138.2. Otherwise, it is the same database. Qiime2 classifiers from the official page were trained before Silva 138.2 was released.
That part puzzles me, but it is probably related to the differences in performance between full-length and region-specific classifiers. Anyway, I usually don't draw any conclusions based on species annotations received with 16S rRNA gene amplicons targeting only 1 or 2 regions.