Defenses & evals: Unify chat templating

[tomtseng]

Summary

Unify all eval prompt formatting on tokenizer.apply_chat_template() by making the template choice an explicit config variable (template_name on ModelConfig).

Motivation

The codebase had a few different formatting mechanisms: (1) manual f-string formatting with user_prefix/assistant_prefix/end_turn, (2) tokenizer.apply_chat_template() using HF built-in Jinja2 templates. This has the risk of being error-prone since there a bunch of different code paths doing the same thing. This PR makes the template choice an explicit config variable and routes all formatting through apply_chat_template().

Child PR #97 refactors attack training data formatting and removes the now-redundant user_prefix/assistant_prefix/end_turn fields from ModelConfig.

Code changes

• Add template_name field to ModelConfig with four options: native (use tokenizer built-in template), generic_chat, instruction_response, plain
• Generate Jinja2 templates from the existing TextTemplate prefix/suffix registry via TextTemplate.to_jinja2()
• Centralize template configuration in load_tokenizer() via configure_tokenizer_template() so evals don't each need their own formatting logic
• Move format_chat_prompt and apply_chat_template_with_fallback from evals/utils.py to whitebox/utils/models/chat_format.py — I hit a circular import otherwise
• Remove model-specific templates (llama3, qwen, gpt_chat) in favor of native — better to just use the template given by the model tokenizer, rather than rewriting it ourselves and potentially getting it wrong

Behavior changes

• StrongReject: Previously received raw (unformatted) prompts. Now prompts go through format_chat_prompt(). With template: plain this is a no-op. With other templates, prompts now get chat formatting -- this will change eval results
• SafetyGap: Bug fix -- the paper code does use chat templating even though our implementation said otherwise.

Child PR: #97

[tomtseng]

this was missing <|begin_of_text|> at the start of a formatted conversation. If we want to apply this kind of formatting I think we should just specify a model name and take its tokenizer.chat_template so we are sure it matches exactly