added support for transcoders to testing and steering endpoints

[curt-tigges]

Problem

We did not support activation testing with transcoders, due to differences in their hook points.

Fix

The SAE manager and "saelens" wrapper have been updated to load transcoders from SAELens, handle their hook points correctly, and use them in activation testing. It is also possible to "steer" with them, but this is currently invalid/undefined and steering should be turned off on transcoders until we figure out what kind of steering would be meaningful.

Testing

I tested this on known features for which we have dashboards, confirming that POST requests to the inference server gave us the expected values. E.g.:

curl -X POST http://127.0.0.1:5002/v1/activation/single \
-H "Content-Type: application/json" \
-d '{
 "prompt": "I love dogs!",
 "model": "gemma-2-2b",
 "source": "0-gemmascope-transcoder-16k",
 "index": "3"
}'

This should show a high activation on the word "love."

[vercel]

Someone is attempting to deploy a commit to the Neuronpedia Team on Vercel.

A member of the Team first needs to authorize it.

[hijohnnylin]

Thanks! @curt-tigges this is blocked on transcoders being merged into SAELens right?

[hijohnnylin]

Moving to draft until we are unblocked on SAELens (which should be soon): decoderesearch/SAELens#182