chat : add the MiniCPM5-2B chat template

[cyxu0401]

Follow-up to #23384 (MiniCPM5-1B), now that
openbmb/MiniCPM5-2B is public.

Adds models/templates/openbmb-MiniCPM5-2B.jinja (fetched verbatim from the
HF repo) and covers it in tests/test-chat.cpp.

2B's chat template is not the same as 1B's. The tool-call and reasoning
syntax is identical, so 2B is already picked up correctly by the existing
detection in common/chat.cpp and resolves to the same chat format. What
differs is how past assistant turns are rendered:

• 1B drops the reasoning content of previous assistant turns entirely;
• 2B keeps it, and when a past turn carries no reasoning it emits an empty
  <think>\n\n</think> block instead.

Nothing covered either behaviour. Since the format is selected by substring
matching on the template text, it is easy to break by accident, so the tests
assert both halves: that 2B still resolves to the shared format, and that the
generation prompt diverges from 1B in the way described above. A negative
assertion was added on the 1B side so the two cannot silently converge.

This is test-only coverage plus a fixture — no runtime code is touched, and
models/templates/ is not read at runtime.

Independent of the tokenizer-note PR; either can land first.

Testing

test-chat passes. I checked the new assertions actually execute by
temporarily inverting one and confirming it fails rather than passing
vacuously.

[ggml-gh-bot]

Hi @cyxu0401, thanks for your contribution!

Per our contribution guidelines, the automated PR checker found the following issue(s) that need your attention:

• PR Template not respected: Please respect the template when creating a new pull request. Make sure to fill out all required sections.

• Multiple open PRs from a new contributor: We limit new contributors (those without a previously merged PR) to 1 open PR at a time. You currently have 2 open PRs.

• AI-generated content: While code is allowed to be generated by AI, please write the PR description and commit messages on your own without the help of AI.

---

Please note that maintainers reserve the right to make final decisions on PRs. If you believe there is a mistake, please comment below.

[cyxu0401]

Closing this. MiniCPM5-2B works with llama.cpp as-is: its tokenizer is byte-identical to 1B, and its chat template ships inside the GGUF, so no upstream change is required. This was only regression coverage, not a fix, and it isn't worth maintainer time right now.