user invites: backend: use randomized repo sampling, not consecutive, for collaborator discovery

Created by: slimsag

Thinking about #31277 (closed) I realize that we're currently choosing the first 20 repos (alphabetically) to discover collaborators. We should probably do a random sampling instead, because consecutive is likely to give similar collaborators (for example, we may hit lsif-java, lsif-typescript, lsif-... all with similar collaborators whereas with random sampling we might hit say the sourcegraph/sourcegraph repo.)