support async mtp

[Ronald1995]

What this PR does / why we need it?

this pr aims to support async_scheduling for mtp, which refer to vllm pr vllm-project/vllm#24799.
and this pr fix some synchronize problem in vllm-ascend.

Does this PR introduce any user-facing change?

How was this patch tested?

• vLLM version: v0.12.0
• vLLM main: https://github.com/vllm-project/vllm/commit/v0.12.0

[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 introduces support for asynchronous Multi-Token Prediction (MTP), which is a significant feature for improving performance with speculative decoding. The changes involve modifications to the rejection sampler, MTP proposer, and the v1 model runner to handle the asynchronous state updates required for this feature. The use of .pin_memory() and other optimizations are good performance improvements.

Additionally, this PR includes a substantial refactoring by adding a new v2 worker implementation (vllm_ascend/worker/v2/). This new implementation seems to align with the upstream vLLM v2 worker API, which is a positive direction for maintainability.

However, the new v2 worker files (aclgraph_utils.py, async_utils.py, model_runner.py) contain several critical issues. They appear to be partially copy-pasted from CUDA-specific code and use torch.cuda APIs instead of the required torch.npu APIs for Ascend devices. This will lead to runtime errors. There is also a bug in the kv_cache_dtype initialization in the new NPUModelRunner. These issues must be addressed before this PR can be merged.

[github-actions]

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