CI: change SGLang plugin M3 nightly accuracy threshold

[Phi-C]

Motivation

For nightly gsm8k accuracy validation, its configuration is

lm_eval \
  --model local-completions \
  --model_args "model=amd/MiniMax-M3-MXFP4,base_url=http://127.0.0.1:8000/v1/completions,num_concurrent=65,max_retries=1,tokenized_requests=False,trust_remote_code=True" \
  --tasks gsm8k \
  --num_fewshot 5

which is different from configurations in recipe

BS=65

lm_eval \
    --model local-chat-completions \
    --model_args "model=${model_path},base_url=http://127.0.0.1:${PORT}/v1/chat/completions,num_concurrent=32,max_gen_toks=16384" \
    --tasks gsm8k \
    --num_fewshot 5 \
    --batch_size "${BS}" \
    --apply_chat_template \
    --fewshot_as_multiturn 2>&1 | tee minimax-m3-mxfp4-sglang-gsm8k.log

So, the accuracy threshold is changed to 0.88.

Technical Details

Test Plan

Test Result

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.

[zhuyuhua-v]

please refer atom's m3 acc test command and threshold, instead of lowering ours..

[Phi-C]

please refer atom's m3 acc test command and threshold, instead of lowering ours..

Okay, the test command is aligned with atom's m3 acc test now.