Add {% generation %} markers for Cohere2 chat template

[qgallouedec]

see #5471

---

Note

Medium Risk
Moderate risk because it changes template auto-swapping logic for Cohere2 and adds a new Jinja template that affects how prompts are rendered during SFT/GRPO, which could subtly impact tokenization and loss masking for those models.

Overview
Adds Cohere2 to TRL’s list of recognized chat-template families and introduces cohere2.jinja plus a patched cohere2_training.jinja that wraps assistant output in {% generation %} / {% endgeneration %} (and moves <|END_OF_TURN_TOKEN|> into role branches) to enable correct assistant_only_loss masking.

Updates get_training_chat_template() to detect the Cohere2 original template and swap in the new training variant, and extends the chat-template test matrix to cover the Cohere2 tokenizer; docs/README are updated to document Cohere2 support and the patch rationale.

^{Reviewed by Cursor Bugbot for commit 0d9bcc0. Bugbot is set up for automated code reviews on this repo. Configure here.}

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: a115d9df13

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[qgallouedec]

merging this with no review: it's been opened for a while an nothing special/potentially breaking