SAE feature coactivation analysis

[JuiceB0xC0de]

Yo I've been messing around with SAE atlasing for a minute and figured you might want some extra action for the atlas.

[bryceag11]

Thanks for adding this! Reviewing today and will pull shortly

[bryceag11]

Co-firing fills a gap, and I ran it on SmolVLA / X-VLA / Pi0.5 / OpenVLA-OFT & it runs clean across all four! Just some cleanups:

• load_sae_checkpoint is a verbatim copy of the one in concept_id.py, import it instead of duplicating.
• Drop the unused get_concept_task_mapping import
• co_fire_rate is the partner's marginal fire rate, not the joint rate, id consider renaming
• Line 137 throws a divide-by-zero warning on sparse models (pi05/oft) so use np.divide(inter, union, out=np.zeros_like(inter), where=union > 0).

Optional: raw Jaccard gets dominated by always-on features (one feature topped 4 unrelated concepts on Pi0.5) so a lift/PMI metric or base-rate filter would sharpen it. And a one-line experiments/README.md entry would be great!

Looks good after those changes ill merge