[Bugfix] Add embed_multimodal to MiniCPM-o 4.5 omni LLM class

[Hiro208]

MiniCPMO45OmniLLMForConditionalGeneration only implements the old
get_multimodal_embeddings. vLLM >= 0.28 calls embed_multimodal directly in V1
encoder-cache profiling and the runtime encoder paths, so the call falls through to
the SupportsMultiModal protocol stub, returns None, and startup fails later in a
NoneType error far from the cause. #7360 has the details.

The sibling wrapper in the same directory (minicpmo_4_5_omni.py) already delegates
embed_multimodal to get_multimodal_embeddings; this adds the same three lines to
the LLM class.

The new test builds the class with object.__new__ (no weights), patches
get_multimodal_embeddings, and checks that embed_multimodal returns its result and
forwards the kwargs.

Verified on CPU with vLLM 0.29.0:

• with the fix, pytest -m "core_model and cpu" tests/model_executor/models/minicpmo_4_5/
  is 423 passed, 1 skipped, and ruff 0.14.10 check/format are clean
• reverting only minicpmo_4_5_omni_llm.py and keeping the new test gives 1 failed
  with assert None is (...): the protocol stub's None, i.e. the bug this closes

Closes #7360

[vllm-omni-review-bot]

This PR appears to belong to: docs/design/module/model_integration.md, docs/design/module/ar_runtime.md.

Module owners: @tzhouam @fake0fan @Gaohan123

Routing: @tzhouam via module of the changed files, CODEOWNERS; @fake0fan via module of the changed files; @Gaohan123 via module of the changed files

@Hiro208, please review your own changes and leave a short self-review comment describing what you checked. PRs without author self-review may not be assigned a reviewer.

Please take a look when you have a chance. If you would like an automated review, mention @vllm-omni-review-bot in a comment.

[Hiro208]

Self-review:

• MiniCPMO45OmniLLMForConditionalGeneration had no embed_multimodal at all (zero
  occurrences in minicpmo_4_5_omni_llm.py on main), so the call resolved to the
  SupportsMultiModal protocol stub and returned None.
• The new method is the same as the one in the sibling minicpmo_4_5_omni.py, docstring
  included, and sits right after get_multimodal_embeddings. That method already returns
  MultiModalEmbeddings, so delegating meets the new API's contract without touching the
  embedding logic.
• The test builds the class with object.__new__, as the cosyvoice3 tests do, so it
  needs no weights and runs in the core_model and cpu lane.
• On vLLM 0.29.0 the minicpmo_4_5 CPU tests are 423 passed / 1 skipped with the fix; with
  minicpmo_4_5_omni_llm.py reverted, the new test fails on assert None is (...).
  ruff 0.14.10 check and format are clean.

[Hiro208]

@vllm-omni-review-bot

[congw729]

LGTM. Verified the #7360 startup fix on a single H200 with real MiniCPM-o 4.5 weights: removing this override reproduces the V1 encoder-profiling failure; keeping it allows profiling, startup, and an image request to complete.

The MiniCPM CPU suite passed (423 passed, 1 skipped, 18 deselected), and all reported CI checks are green, including Buildkite #15005 (21 passed, 0 failed).

The bare-LLM image probe still produced repetitive text; output correctness remains a separate follow-up and is not covered by this startup-fix validation.

AI-assisted review and validation with Codex.