Extracting row embeddings from TabFM for similarity search

Hi, thanks for releasing TabFM!

I started an experiment today using TabFM as a row embedding model for similarity search over tables: fit on a context table, embed held-out rows, retrieve nearest neighbors by cosine similarity. I'm extracting the output of `row_interactor_2` (2048-d) with a forward hook, and comparing it to TabICL and TabPFN behind one API: https://github.com/vstenby/tfm-embeddings

It's day one and I'm still figuring out if this is a useful direction, so early feedback is welcome. Specific questions:

1. Is `row_interactor_2` where you'd extract a general-purpose row representation? With two column-embed/row-interact stages I wasn't sure if the second stage is more task-specialized and the first might transfer better.
2. Would you consider exposing row representations as a supported API? Happy to contribute if there's interest.

All the best,
Viktor

Start with the row_interactor_2 output and the forward-hook experiment described in the linked tfm-embeddings project. Compare the first and second row-interact stages, then confirm with maintainers whether either representation should become a supported API. No files, tests, or acceptance criteria are named, so the desired public interface and definition of done need clarification.

Written by the indexing model from the issue text.