[Performance]use triton mrope for Qwen3-VL

[ichaoren]

What this PR does / why we need it?

The performance can be optimized by use triton mrope for Qwen3-VL.

1. Qwen3-VL-235B-A22B-Instruct-W8A8

	accuracy	TTFT	TPOT	TPS
base	83.76	4.8771s	0.1472	49
test	83.59	4.3273	0.0615	105

2. Qwen3-VL-8B-Instruct

	accuracy	TTFT	TPOT	TPS
base	80.65	4.1744	0.0499	125
test	80.86	3.1858	0.0245	227

Does this PR introduce any user-facing change?

Name: triton-ascend
Version: 3.2.0.dev20260105

How was this patch tested?

vllm serve $MODEL_DIR
--served-model-name $MODEL_NAME
--quantization ascend
--tensor-parallel-size 8
--data-parallel-size 1
--enable-expert-parallel
--max-model-len 20000
--allowed-local-media-path /vllm-workspace/vllm/tests/multimodal/assets/
--enforce-eager
--max-num-batched-tokens 5000
--gpu-memory-utilization 0.8
--trust-remote-code
--port 8800
--no-enable-prefix-caching
--mm_processor_cache_type="shm"
--async-scheduling

• performance test
  evalscope perf --parallel 8 --model qwen3vl --url http://ip:port/v1/chat/completions --api openai --dataset random_vl --min-tokens 300 --max-tokens 300 --prefix-length 0 --min-prompt-length 1024 --max-prompt-length 1024 --image-width 2048 --image-height 2048 --image-format RGB --image-num 1 --number 32 --tokenizer-path /path/Qwen3-VL-235B-A22B-Instruct-W8A8/

• Accuracy Test
  ais_bench --models vllm_api_stream_chat --datasets textvqa_gen --debug

[gemini-code-assist]

Code Review

This pull request introduces a performance optimization for Qwen3-VL models by enabling a Triton-based multi-head rotary position embedding (mrope) implementation. The change correctly identifies the specific model configuration and directs it to a more efficient code path. My review includes one suggestion to improve code maintainability by replacing a hardcoded value with a named constant.