Compute activations other than MLP neurons

[tim-lawson]

FYI: I don't know whether you're open to contributions, but someone might find this helpful.

Modify the activations lib and expgen project to generate exemplars for activations other than MLP neurons -- the residual stream, MLP in/outs, and self-attention outs. These match the Subject collect_acts method, excluding self-attention maps (different shape) and unembed in/outs (no layer index).

Follows the recommendation in project/expgen/README.md by adding to get_activations_computing_func -- specifically, introduces an ActivationType enum which determines access paths for subject -> component and component -> activations.

Adds an activation-type suffix to the exemplars folder name, except when the activation type is "neurons," to preserve current behavior. Similarly, adds a command-line argument to the expgen compute_exemplars.py script with default value "neurons".

Caveat: not tested with the rest of the expgen pipeline yet.

[choidami]

Thank you for implementing this! We are totally open to contributions (it's precisely why we open sourced our code!).
I'll merge the changes in once I test the integration with the expgen pipeline.

[tim-lawson]

You're welcome! Thanks for open-sourcing it. There's no rush to merge; if you find there are incompatibilities with the pipeline as-is, I can revisit the changes.