Backtracking is coarser than necessary

Analysis of the `AssignmentSet` changelogs seems to indicate that backtracking in the sat-solver is far too aggressive. What typically happens is a whole suite of packages are suggested by the policy and assumed `True` until finally a conflict is found. Then _everything_ is unwound back to the beginning and almost all of the same packages are re-suggested by the policy. This repeats many times over the course of the search.

To fix this, we need to be able to backtrack to the previous suggested package and not back to the root of the search tree. I think it's _supposed_ to work this way already, so closing this will either require a bug fix or a good explanation for why it's happening.