Draft: Rebuilderd stats speed up

Instead of running one query per error category, run one query per BAD package and then process the errors Python-side, short-circuiting the matches.

The processing time on the i386 database goes from 30 seconds to 11 seconds.

A handful packages are classified differently (but I'd argue rightly so), and the catch-all category "failed to reproduce" must now be at the end (can be fixed later in the presentation phase).