feat: cluster derivations by package

[adekoder]

This PR is the Phase 1 of Nix Derivation data De-duplication.

The homepage and description are duplicate across evaluation of the same packages, we are extracting them into
NixDerivationHomepage and NixDerivationDescription tables as a first step toward package-level deduplication.

Notes:

1. Run manage.py migrate_homepage_description --dry-run — verify counts
2. Run manage.py migrate_homepage_description
3. Run manage.py migrate_homepage_description --null-after to clear the old columns

[fricklerhandwerk]

Only reviewed the tests, let's fix those first

implemented

[fricklerhandwerk]

Rebased and reordered the commits to have the test refactoring be a preparation rather than an afterthought.

[fricklerhandwerk]

Rebased on top of #537, since we decided to just go ahead and do the package clustering in this PR already. Reason: We'd be moving >100M rows, which we'll have to go through again if we clustered in a follow-up. We're already wasting disk space, and may as well deduplicate immediately.

[fricklerhandwerk]

I've implemented package clustering as discussed with @adekoder:

• on trigger after complete eval
• as a command
• tests
• removed the previous merely-moving of data

There were few interesting edge cases to consider that weren't obvious so far, such as how to organize the catch-up so it doesn't clobber a fresh eval if it's still running.

Closes #790 now.

Marked as draft, since there are still somewhat unrelated preliminary changes baked in. I split them out into separate PRs, after which I'll rebase this one:

• test: send notification explicitly #1049
• perf: prefetch channel for cached suggestions #1050
• refactor: move rolling-release predicate to NixChannel #1051
• refactor: move descriptions search vector to package data model #1052

[fricklerhandwerk]

Reviewed with @adekoder, who correctly pointed out the remaining race condition between the post-eval listener and the backfill.

• Renamed the command to highlight that it's for backfilling

  It now starts with the services to recover what crashed listeners may have left over

• Cut off the backfill query up to the latest complete evaluation.

  This ensures that listener and backfill command never touch the same rows

• Wrapped each batch into a transaction to avoid partial updates (especially deleting the metadata from the source derivation!)

  @adekoder Probably we don't need row locking after all (it's still there after the messy fix), since the concurrent queries are now non-overlapping

Left to do:

• Review tests
• Test the concurrency scenario
• Resolve merge conflicts
• Massage history

[fricklerhandwerk]

skip_locked=True only prevents concurrent access to the same derivation, but not Package, PackageAttrpath, or the overall clustering decision.

If backfill and post-eval trigger race at the batch boundary, we can end up with

• attrpath foo.bar registered to package P1
• some derivations with foo.bar linked to P1
• other derivations with foo.bar linked to P2

This will produce cleanup work and, likely, extreme confusion. I prefer fixing now this since we don't have UI for cleanups, and it will take a while until we get to that UI. Working on a draft.