Add Direct Logit Attribution tool for TransformerBridge

[TravisHaa]

Description

• Implemented a Direct Logit Attribution (DLA) tool for the new TransformerBridge system, closes [Proposal] Direct Logit Attribution Tool #1263. Based on the stale PR (Draft) Add DLA function to utils #466 but adapted to utilize 3.0 TransformerBridge.
• returns per-component (or per-layer) contributions to logit difference ebtween correct and wrong token, decomposing residual stream based off accumulated bool.
• generated docstring (according to contributing.md) highlighting important warnings of limitations of DLA tool (currently does not support hybrid layers like mamba, requires bridge compatibility mode)

Acknowledged limitations

• Strict mode only. hybrid architectures (Mamba, SSM, Mixer, LinearAttention) raise
  NotImplementedError. ActivationCache.decompose_resid only knows how to decompose attn_out + mlp_out per layer; supporting hybrid blocks requires extending that method and is out of scope
  for this PR. (will be working on this in next steps)
• Requires compatibility mode. Raises ValueError if bridge.enable_compatibility_mode()
  hasn't been called. Without folded LayerNorm weights, the projection direction is wrong and
  per-component scores don't reflect actual logit contributions.
• Excludes unembedding bias. Per-component scores sum to actual_logit_diff − (b_U[correct] − b_U[wrong]). This matches the convention in cache.decompose_resid and PR (Draft) Add DLA function to utils #466 — the bias is a constant offset

Type of change

• New feature (non-breaking change which adds functionality)
• This change requires a documentation update

Screenshots

Checklist:

• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes
• I have not rewritten tests relating to key interfaces which would affect backward compatibility

[jlarson4]

Hi @TravisHaa! I have reviewed your PR and left a few comments. I cannot run the code in its current state, until some of these comments are addressed. Let me know if you have any questions.

Also, the PR is marked as "tests added", but I am not seeing any tests in the diff. Could you add tests/unit/tools/test_direct_logit_attribution.py covering at least:

• Correct path on a small bridge (e.g., gpt2 with compatibility mode)
• ValueError when compatibility mode is off
• NotImplementedError when a Mamba-like adapter is present
• Both accumulated=True and accumulated=False

Feel free to tag me once these edits are in and I will re-review. Thank you for your work on this, it is coming along nicely!

[TravisHaa]

@jlarson4 I apologize for the delay once again, I spent a lot of time understanding how I can improve this PR on top of your suggested fixes by exploring the codebase. I will make sure to re-iterate within a couple hours of your review for this brand new commit 👍