Robustness: Having to essentially play back a log to recover any particular revision increases a chance of something eventually getting corrupted, and so it's somewhat safer to avoid it in something intended to be archival.
Ease of extracting subsets: For researchers, having the revisions be independent allows you to filter the XML dump through a SAX parser (or similar) to grab only revisions meeting particular criteria. If deltas were stored, you'd have to reconstruct those revisions from the deltas, which would make it really expensive to do things like, "I want to look at every article as it appeared at noon on April 1, 2007".
In the live DB, I think just because it's cheaper to get a ton of storage, esp. for rarely-retrieved old revisions, than to add the overhead of computing deltas and applying them to reconstruct revisions. In particular, you'd have to compute a diff for every edit in that situation, whereas currently MediaWiki only computes diffs when a user requests to view one from the "history" tab, which is a tiny proportion of all edits.
IIRC from some tests a year or so ago, Wikipedia hasn't found any significant improvement from 7z over bz2 on the current-revisions-only dump, which looks more like just normal English text; that's why it doesn't bother to provide a separate 7z version of that. It seems to only be this pattern of [200kb article][almost the same 200kb article][almost the same 200kb article again] that 7z kills bz2 on.