Refactor flashinfer logic for deepseek v3 and fix accuracy bug

[Fridge003]

Motivation

flashinfer_backend.py for attention is too complex, this PR extract the logic of MLA and creates a new flashinfer_mla_backend.py

Also, #3716 #3751 reports an accuracy bug when enabling flashinfer mla. This PR solves this bug by correctly handling rope scaling with yarn, with the help of @yzh119.

Modifications

• Define FlashInferMLAAttnBackend in flashinfer_mla_backend.py by removing codes irrelevant to MLA in flashinfer_backend.py
• Remove magic numbers in code so Deepseek v2 can also be supported
• Simplify the code in forward of MLA
• Change flash attention backend to auto so the newest fa3 backend of flashinfer can be used
• Fix accuracy bug with yarn rope scaling

Accuracy Test

The baseline results of not enabling flashinfer mla can be referred to #3486.

Server

python3 -m sglang.launch_server --model deepseek-ai/DeepSeek-R1 --tp 8 --trust-remote-code --enable-flashinfer-mla

gsm8k

python3 benchmark/gsm8k/bench_sglang.py --num-shots 8 --num-questions 1319 --parallel 1319

Accuracy: 0.960
Invalid: 0.000
Latency: 101.807 s
Output throughput: 1329.111 token/s

mmlu

bash benchmark/mmlu/download_data.sh
python3 benchmark/mmlu/bench_sglang.py --nsub 100 --ntrain 5 --parallel 2000

Total latency: 159.003
Average accuracy: 0.871

Checklist

• Format your code according to the Code Formatting with Pre-Commit.
• Add unit tests as outlined in the Running Unit Tests.
• Update documentation / docstrings / example tutorials as needed, according to Writing Documentation.
• Provide throughput / latency benchmark results and accuracy evaluation results as needed, according to Benchmark and Profiling and Accuracy Results.
• For reviewers: If you haven't made any contributions to this PR and are only assisting with merging the main branch, please remove yourself as a co-author when merging the PR.
• Please feel free to join our Slack channel at https://slack.sglang.ai to discuss your PR.