Molmo2 tokenizer chat template rejects system turns (roles must alternate user/assistant)

**Affects:** the `allenai/Molmo2-4B`/`8B` HF tokenizers' built-in chat template (relevant to the `vision` branch's SFT pipeline and any consumer tokenizing standard SFT data with the tokenizer-native template).

Applying the Molmo2 tokenizer's built-in chat template to conversations containing a system message raises:

```
jinja2.exceptions.TemplateError: Conversation roles must alternate user/assistant/user/assistant/...
```

mm_olmo handled system turns by flattening them into the first user message (`olmo/data/tulu.py`-style `"System: {content}\n{user}"`), but the exported HF template enforces strict user/assistant alternation instead — so real SFT mixes (e.g. `allenai/Dolci-Instruct-SFT`, where many rows carry system turns) cannot be tokenized with the tokenizer-native template.

**Suggested fix:** make the template accept a system role (render it as a proper system turn, or fold it into the first user turn to match mm_olmo training behavior).

Found while mixing text SFT data into stage-2 training in allenai/open-instruct (open-instruct#1856); we work around it with a registered system-capable ChatML template.