[Attention][Bugfix] Fix FA sink support

[MatthewBonanni]

Purpose

#24794 didn't mark FlashAttention as supporting sink, so GPT-OSS ends up selecting TRITON_ATTN, causing failure. This PR fixes the issue.

Test Plan

python examples/offline_inference/spec_decode.py \ --method "eagle3" \ --tp 1 \ --print-output \ --model-dir openai/gpt-oss-20b" \ --eagle-dir "RedHatAI/gpt-oss-20b-speculator.eagle3" \ --dataset_name "hf" \ --dataset_path "philschmid/mt-bench" \ --num-spec-tokens 3 

Test Result

Previously failed, now works

---

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.

[gemini-code-assist]

Code Review

This pull request correctly addresses the bug where FlashAttention was not marked as supporting sink tokens, leading to incorrect backend selection for GPT-OSS models. The added supports_sink class method in FlashAttentionBackend properly checks for the availability of FlashAttention's variable-length function and its sink support. The implementation is straightforward and directly resolves the reported issue, ensuring that the correct attention backend is utilized.