Built in Logit Lens

[jlarson4]

Description

Created a simple tool for doing basic logit analysis. Being able to read the residual stream after each individual layer is a common research pattern that did not yet have a direct analog in TransformerLens, this intends to close that gap.

The PR adds a single-call tool to transformer_lens.tools.analysis:

• logit_lens(model, prompt) or cache.logit_lens() reads each entry of the accumulated residual stream through the model's final norm and unembedding and returns per-layer logits, log-probabilities, target-token ranks, top-k tokens or a vocabulary subset.
• logit_readout(model, vectors) does the equivalent for any d_model tensor (head outputs, steering directions, sparse-coder decoder columns).
• targets can be a [batch, pos] tensor of next-token ids, producing the rank of the next token at every layer & position.
• Results carry rank_trajectory, top_tokens, decode and to_dataframe.
• Fixes Cohere in compatibility mode, where the adapter applied logit_scale twice after folding it into the unembedding.
• Docs section in analysis_tools.md; 32 unit, 12 integration and 2 adapter tests.

Type of change

• New feature (non-breaking change which adds functionality)
• Bug Fix
• Documentation Update

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