Improve task scheduler to not require model as an axis

[jjallaire]

The first type of task parallelization we implemented was running a task against n models. We later added more arbitrary task parallelization however in doing so remained over-indexed on the model axis. We need to re-write our scheduler to take more arbitrary combinations of task/model.

This will in turn make eval-set scheduling more effecient (as retries won't need work in "passes" based on model permutations).