Revisit dependency solving strategy

[dralley]

Our goal is of course to remove dependency solving to the greatest extent possible. However, if it's not possible to remove it completely, there are potentially other models that are better suited to the process (enable more performance, etc.) than libsolv.

1. We could perform dependency solving separately from the initial copies, in order to avoid needing to do it more than once due to request size limits.

2. We could switch to an entirely different model, such as PubGrub, or Resolvo that is not reliant on complete satisfiability. There are some libraries available for this and it could have a dramatic impact on performance and other metrics - potentially.