refact model runner v1

[weiguihua2]

What this PR does / why we need it?

1. Remove the init torchair logic from model_runner_v1 and put it into torchair_model_runner
2. Separate the execute model logic from the prepare input logic

Does this PR introduce any user-facing change?

How was this patch tested?

• vLLM version: v0.10.0
• vLLM main: vllm-project/vllm@fda9537

[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.

[github-actions]

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

[gemini-code-assist]

Code Review

This pull request is a significant and well-executed refactoring of the model runner logic. The key changes, including separating the torchair-specific logic into NPUTorchairModelRunner and splitting the input preparation from model execution in NPUModelRunner, greatly improve the code's structure, clarity, and maintainability. The introduction of AscendCommonAttentionMetadata is also a good step towards better-structured code.

I've found one critical issue where the _generate_process_reqs_hidden_states method in the base NPUModelRunner is missing required arguments in its call to the model's forward method. This would likely cause a TypeError in non-torchair execution paths. Please see the detailed comment for the fix.

[gemini-code-assist]

The call to self.model() is missing the kv_caches and attn_metadata arguments. The model's forward method requires these arguments to be passed. While they are available in the forward context for attention layers, the top-level model forward signature still needs them. This will likely cause a TypeError at runtime for non-torchair execution paths.

Suggested change

	hidden_states = self.model(
	input_ids=input_ids,
	positions=positions,
	intermediate_tensors=intermediate_tensors,
	inputs_embeds=inputs_embeds,
	)
	hidden_states = self.model(
	input_ids=input_ids,
	positions=positions,
	kv_caches=self.kv_caches,
	attn_metadata=attn_metadata,
	intermediate_tensors=intermediate_tensors,
	inputs_embeds=inputs_embeds,
	)