[Feature] Support flashinfer allreduce fusion tuning

[BBuf]

Checklist

• 1. If the issue you raised is not a feature but a question, please raise a discussion at https://github.com/sgl-project/sglang/discussions/new/choose Otherwise, it will be closed.
• 2. Please use English, otherwise it will be closed.

Motivation

We need tuning flashinfer allreduce fusion workspace buffer sizes to get best performance in more general cases. Currently, we hardcoded max_token_num=2048 at here .

Refer to benchmark results here .

Related resources

No response

[github-actions]

This issue has been automatically closed due to inactivity. Please feel free to reopen it if needed.