Current status: we query all tags sequentially, and then merge packages into one big package set, from which we generate one big repo for each architecture. For each variant we track which tags it should use.

Proposal: We should refactor this to create one set of repos for each tag. This would provide better options for runing stuff in parallel, plus we would no longer need the package whitelist hack in gathering.

The package set dict in memory should probably remain as one big structure for now.