[Refactor] Separate _prepare_inputs to _prepare_inputs and _preprocess

[gcanlin]

What this PR does / why we need it?

Part of RFC #5449.

Align with upstream vLLM. This PR will help downstream vLLM-Omni reduce the cost for maintaining the _prepare_inputs. Besides, it helps vLLM-Ascend code more readable. In the future, we can follow closer to vLLM.

• Moved the multimodal, prompt-embed, positions, PP handling and update_cos_sin into _preprocess, and trimmed _prepare_inputs to return only metadata plus logits and spec-decode inputs.
• Updated execute_model to call _prepare_inputs then _preprocess, preserving the original ordering while separating concerns.
• Reuse _prepare_mm_inputs in vLLM and add model_kwargs.

NOTE: This PR includes #5971 changes. We need to wait it merged(if it would be approved). Then rebase this PR.

Does this PR introduce any user-facing change?

How was this patch tested?

• vLLM version: v0.13.0
• vLLM main: vllm-project/vllm@2c24bc6

[github-actions]

👋 Hi! Thank you for contributing to the vLLM Ascend project. The following points will speed up your PR merge:‌‌

• A PR should do only one thing, smaller PRs enable faster reviews.
• Every PR should include unit tests and end-to-end tests ‌to ensure it works and is not broken by other future PRs.
• Write the commit message by fulfilling the PR description to help reviewer and future developers understand.

If CI fails, you can run linting and testing checks locally according Contributing and Testing.

[gemini-code-assist]

Code Review

This pull request refactors the _prepare_inputs method in NPUModelRunner to better align with the upstream vLLM implementation by creating a new _preprocess method. Logic for multimodal inputs, prompt embeddings, positions, and pipeline parallelism has been moved from _prepare_inputs to _preprocess. Consequently, execute_model is updated to call these two methods sequentially, which improves separation of concerns while preserving the original execution order. The refactoring is clean, simplifies data flow by removing maybe_padded_num_tokens, and updates method signatures consistently. This is a solid improvement for maintainability and alignment with upstream. I have not found any issues with the changes.

[gcanlin]

@wangxiyuan Could you please help add the ready tag for e2e-full test? I want to run it in the free weekend to avoid resource queuing. Thanks!

[gcanlin]

@zhenwenqi2024 Hi! Could you please help review this PR? Thx!

[gcanlin]

Qwen-Omni needs mm_kwargs

[github-actions]

This pull request has conflicts, please resolve those before we can evaluate the pull request.