[Feature] add kernel level benchmark

[zhyncs]

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

use triton benchmark utils https://triton-lang.org/main/python-api/generated/triton.testing.do_bench.html#triton.testing.do_bench to benchmark kernels (flashinfer, triton, vllm, tensorrt llm, cudnn etc)

Related resources

No response

[zhyncs]

Currently, the main branch already has some e2e benchmarks, but this is far from enough for us. We need kernel-level benchmarks, ranging from small batch size to large batch size, various shapes, and data types. Compared to existing libraries, what level are we at, and how much room for optimization is there?

[zhyncs]

ref #2486