Allow aggregate jobs to be created from a list of job names

[edolstra]

Evaluating an aggregate job takes as much memory as the sum of its constituents - no constituent expression can be garbage-collected while the top-level expression is reachable. For example, the tested job in NixOS's release-combined.nix takes > 12 GB.

As an alternative, we could support aggregates to be specified as a list of job names, e.g.

{
  tested = {
    type = "hydraAggregate";
    constituents =
      [ "nixos.iso_plasma5.x86_64-linux"
        "nixos.iso_minimal.aarch64-linux"
        ...
      ];
  };
}

hydra-eval-jobset can then lookup the constituent's in the JSON output of hydra-eval-jobs and create a derivation that depends on the drvPaths of the constituents.

[vcunat]

Well, we do have evaluation now, but the tested jobsets get aborted after all constituents succeeding (example link). For unstable-small it started in the small range 1e0222cd1f6..895042956f2 containing the change, so it seems clear to have triggered the problem.

EDIT: logs in the web UI show nothing, so I can't see how to push this forward myself.

[edolstra]

Looks like we're generating a bad derivation (...drv.drv).

[edolstra]

Fixed in NixOS/nix@edee616.