Fix ChatVLLM crash on models without a chat template

[ferreirafabio]

What is the problem?

PR #8 introduced the ChatVLLM wrapper which switches from vllm.LLM.generate() to vllm.LLM.chat() so that chat templates get applied correctly. This works great for instruct models that ship a chat template in their tokenizer config. However, base/pretrained models like swiss-ai/Apertus-8B-2509 don't define one. Since transformers >= v4.44 no longer provides a default chat template, calling vllm.LLM.chat() on these models raises a ValueError. This also means we can't evaluate base models for fluency tasks anymore, which is something we need for the project.

How do we solve it?

We detect at model load time whether a chat template is available and pick the right vLLM method accordingly. Three paths:

1. User passes --chat_template: we use llm.chat() with that explicit template. Useful when you know the right format for a model whose tokenizer doesn't include one.
2. Tokenizer has a chat template: we use llm.chat() and let vLLM apply it automatically. This is the default path for instruct models.
3. No chat template found: we fall back to llm.generate() (plain text completion, no chat formatting) and print a warning. This is the expected path for base models used in fluency evaluation.

This way instruct models keep working as before, base models no longer crash, and users can still force a template via the CLI when needed.

Changes

• openjury/utils.py: add warnings import, three-path template detection in ChatVLLM.__init__(), new _to_raw_text() method for the generate() fallback, pass chat_template through batch() and make_model()
• openjury/generate.py: forward chat_template parameter in generate_instructions() and generate_base()
• openjury/generate_and_evaluate.py: add --chat_template CLI argument, thread it through CliArgs, gen_fun partials, and make_model() calls
• README.md: document chat template behavior under "Model Specification"

Testing

Tested on L40S GPU with vllm 0.10.2 using both Apertus 8B models:

• swiss-ai/Apertus-8B-2509 (base, no chat template): correctly falls back to llm.generate(), warning emitted, produces valid completions
• swiss-ai/Apertus-8B-Instruct-2509 (instruct, has chat template): correctly uses llm.chat() with the tokenizer's template
• swiss-ai/Apertus-8B-2509 + explicit ChatML template: correctly uses llm.chat() with the provided override
• make_model("VLLM/...") end-to-end: chat_template parameter correctly forwarded through the full pipeline