Add clustering analysis

[danbraunai]

Migrated from: goodfire-ai/spd-gf#51
Original author: @leesharkey

---

SPD produces a set of subcomponents which need to be clustered together in order to identify parameter components. Some of the models in our evals suite will have parameter components that must be clustered together from subcomponents. We will want to evaluate the results of this clustering process in order to test our analysis pipeline end to end. If either SPD or clustering are broken, then we will end up with poor results. So we need to evaluate our clustering method.

It will probably suffice to use the clustering methods on our toy models where we have ground truth. We can usually identify if SPD has succeeded without evaluating the subsequent clustering. But the clustering should still work! So we should test the clustering separately from our SPD metrics.

[danbraunai]

Changed title to "Add clustering analysis", because I'm not sure we should be thinking about merging it with any wandb reports or anything yet. Just having good standalone scripts to run clustering would be a great first step. @mivanit I think we discussed this but just mentioning in case we weren't on the same page here and wanted to talk about it.