[Tracking][Benchmark] Add more workloads in ATOM/Benchmark Actions

[wuhuikx]

We want to add more workloads in the Actions of ATOM/Benchmark.
The models with functionality ready will be kept the corresponding recipes in https://github.com/ROCm/ATOM/tree/main/recipes

• Qwen3-235B FP8, update the recipe with nightly docker @gbyu-amd
• Qwen3-235B FP4, ensure the functionality first then update the recipe @PerryZhang01
• Kimi-K2-Thinking FP4, add to the Benchmark @gbyu-amd
• Qwen3-next FP8, w and w/o MTP, add a recipe after the model ready @ganyi1996ppo

[PerryZhang01]

🔍 CI Enhancement: Model Accuracy Validation

Currently, our CI pipeline only performs functional verification of models without assessing their inference accuracy. It is essential to establish minimum accuracy thresholds for each model on the GSM8K dataset. Below are the proposed accuracy standards and validation requirements:

📊 GSM8K Accuracy Benchmarks

Model	Minimum Accuracy Threshold
Deepseek-R1-0528-FP8-MTP	≥ 94%
Deepseek-R1-0528-FP4-MTP	≥ 93%
Deepseek-R1-0528-FP4	≥ 93%
Deepseek-R1-0528-FP8	≥ 94%
gpt-oss-120b-2gpu	≥ 38%
gpt-oss-120b	≥ 38%
Llama-3.3-70B-FP4	≥ 88%
Llama-3-8B	≥ 73%
Qwen235B-A22B-FP8	≥ 87%
Qwen235B-A22B-FP4	≥ 86%
Qwen3-Next	≥ 85%
Qwen3-Next-MTP	≥ 85%

please @valarLip help fill out the above form and @gyohuangxin help set up the accuracy test.