Add per-class token and sentence attributions with interactive visualizer

[ianbulovic]

Summary

Lots of help from Claude on this one, but carefully steered and scrutinized by me.

• Adds two new interpretation methods to the REST API: gradient-based token attributions and leave-one-out sentence attributions. This is adapted from @tmills rest_interp branch, slightly modified to return signed per-class scores rather than a single unsigned saliency score
• Adds a GET / endpoint that serves an interactive HTML visualizer for exploring model predictions and attributions
• Adds tests for both attribution methods against the existing negation model

Token attributions (return_attributions=true)

Uses the input × gradient method: one forward pass through the encoder, then one backward pass per class label (reusing the computation graph via retain_graph=True). The signed dot product of gradient and embedding over the hidden dimension gives a per-token, per-class score in [-1, 1]:

• Positive: token pushes the model toward that class
• Negative: token pushes the model away from that class

Each token in the response includes token_id, character-level start/end offsets into the original input string, and a scores dict keyed by label. Special tokens ([CLS], [SEP]) are included with start == end == 0. Only classification tasks are supported; tagging and relations tasks log a warning and are skipped.

Sentence attributions (return_sentence_attributions=true)

Uses leave-one-out ablation: runs one batched forward pass over the full text and each ablated variant (one sentence removed), then for each class computes p(class | full) − p(class | ablated). Scores are per-class and signed with the same polarity convention. Single-sentence inputs return a score of 0.0 for all classes.

Visualizer (GET /)

A self-contained HTML page (no external dependencies) served at the root of each model's route prefix. Features:

• Form controls for all /process query parameters
• Classification results with prediction, per-class probability badges
• Token attribution display: original text rendered with colored highlights (green/red by class score intensity), class selector dropdown, hover tooltips showing all class scores
• Sentence attribution display with signed scores and class selector
• Basic span/relation lists for tagging and relation tasks

The HTML file is bundled as package data via importlib.resources and declared in pyproject.toml so it is included correctly in installed wheels.