[Model Runner V2] Support multi-modal embeddings for spec decode model

[TheEpicDolphin]

Purpose

Currently the multimodal inputs from the target model vision encoder are not passed to the draft model. Draft models trained on vision embeddings need to merge these inputs with the input token ids to function properly. This behavior is supported in MRV1, but not yet in MRV2. This PR addresses that gap.

Manual Testing

Verified no regression in acceptance rate for the following vision models:

1. VLLM_USE_V2_MODEL_RUNNER=1 vllm serve Qwen/Qwen2.5-VL-7B-Instruct --no-enable-prefix-caching --tensor-parallel-size=1 --data-parallel-size=1 --speculative-config '{"method": "eagle3", "model": "Rayzl/qwen2.5-vl-7b-eagle3-sgl", "num_speculative_tokens": 3}'

2. VLLM_USE_V2_MODEL_RUNNER=1 vllm serve Qwen/Qwen3-VL-4B-Instruct --no-enable-prefix-caching --tensor-parallel-size=1 --data-parallel-size=1 --speculative-config '{"method": "eagle3", "model": "AngelSlim/Qwen3-VL-4B-Instruct_eagle3", "num_speculative_tokens": 3}'

Unfortunately, I couldn't find any target/draft model pairs that are both trained on vision embeddings and also currently work on MRV2 to test on and observe an increase in acceptance rate. So I instead tested the scenario where the draft model receives vision embeddings by forcing self.supports_mm_inputs to True, and checking that the correct number of embeddings are passed and merged with the input ids during draft model prefill.

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @TheEpicDolphin.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[gemini-code-assist]

Code Review

This pull request adds support for multi-modal embeddings in the speculative decoding model runner. The changes involve propagating scheduler_output to access encoder inputs and preparing multi-modal embeddings for the speculative step. The logic appears correct, but I've identified an opportunity to improve code maintainability by refactoring duplicated code.

[WoosukKwon]

Can you please rebase?

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @TheEpicDolphin.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @TheEpicDolphin.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[WoosukKwon]

LGTM. Thanks for the PR! I've made small edits in the PR for minor code cleanup.