[Perf] refactor attention backend for perf boost

[ganyi1996ppo]

Purpose

The current aiter_fa_backend contains redundant fetch kv operation every run, which may cause not only serious memory pressure, but also the time consumption on every decode layer's attention inference. Based on that observation, we rewrite the attention backend to eliminate the unnecessary fetch kv operation, and rewrite the fetch kv triton kernel for better occupancy when chunk prefill or similar scenario occurs. Also in order to prevent unnecessary memory reorder operation, there are some changes on both model_runner and scheduler, which introduce negligible host overhead as we tested. The detail analysis can be found in following doc https://amd.atlassian.net/wiki/spaces/MLSE/pages/1143506837/AITER+Attention+Backend+Proposal

We have tested the perf on current attention, on Qwen3 for both 30B and 235B. The perf boost looks great, especially comes to the long prompt scenario. Test result also attached into the upper doc.

Test Plan

Test Result

We observe nearly 4.x throughput boost on long prompt short output case, also with significant latency improvement. this is the benchmark result, the test script also attached in the upper doc:
prev

new:

---

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.
• (Optional) Release notes update. If your change is user facing, please update the release notes draft in the Google Doc.

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.

💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.

Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors.

You ask your reviewers to trigger select CI tests on top of fastcheck CI.

Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can either: Add ready label to the PR or enable auto-merge.

If you have any questions, please reach out to us on Slack at https://slack.vllm.ai.

🚀

[ganyi1996ppo]

The PR to vllm is in vllm-project#25763