Track-wise Optimization Options

Some observations from @fxiafxia and others who worked with the reviewer assignment system for ACL-2021:
- Breaking down the optimization problem into subproblems/submatrices for each track _greatly_ improved the overall runtime of the optimization
- This makes sense because it eliminates a huge amount of unused parameters
- We implemented this optimization with our acl-2021 branch
- With the subproblem optimization, the step that takes by far the longest now is the aggregation of paper-paper similarity scores into paper-reviewer similarity scores
- The results of the optimization system seem to be very "unbalanced". I.e. a few reviewers will get the max number of papers and lots of others will get none or the minimum
- It is hard to "tune" the min and max number of papers to assign, since the needs of each track are quite different from each other. Some tracks have an excess of reviewers, some have barely enough

Based on these observations, we had a few thoughts for the future of the system:
- Would it be possible to only compute similarity scores and aggregate scores within tracks, since this step seems to take a lot of time globally? (if the `--track` parameter is specified)
- Could you read in a configuration file that has the min and max papers per reviewer on a track-by-track basis, and defaulting to the global setting when this is not included?
- Could you somehow tweak the objective function to consider "balanced" assignment?
- A naive formulation of this might be adding up the differences between each reviewer's assigned number of papers and the "track average", if you divided the papers equally among the reviewers of the track
- Others have put forward ideas taking this notion of "balance" further, e.g. balancing inexperienced vs experienced reviewers, academia vs industry, geographic balance, etc