That's a big win in terms of disk space, and shouldn't affect performance
negatively (there's only an extra query in `content_add` on hashsha1/sha1_git collisions)
Resolves T2304.
When this diff is accepted, I'll do a similar one for skipped_content.