[feat][sgl-atom] add qwen3 sglang dense model support

[zhangxinyuanliuhengyu]

Summary

• Register Qwen3ForCausalLM in the SGLang+ATOM model adapter registry so Qwen3-32B-FP8 uses the ATOM plugin model path instead of falling back to the built-in SGLang model.
• Set --page-size 16 for the Qwen3-32B-FP8 benchmark entries to avoid the dense attention KV layout issue seen with page_size=1.
• Keep the existing dense attention gating behavior unchanged.

Accuracy Validation

Validated Qwen3-32B-FP8 with the SGLang+ATOM service using the same --page-size 16 runtime configuration.

The initial GSM8K CI-style run looked low because lm_eval local-completions defaults to max_gen_toks=256, which truncates Qwen3 math reasoning outputs. A direct A/B confirmed this was an evaluation configuration issue, not a model accuracy regression:

• GSM8K 3-shot, 100-sample smoke, default max_gen_toks=256: flexible-extract=0.56, strict-match=0.56
• GSM8K 3-shot, 100-sample smoke, max_gen_toks=2048, max_length=8192: flexible-extract=0.95, strict-match=0.97
• GSM8K 3-shot, full set, max_gen_toks=2048, max_length=8192: flexible-extract=0.9037, strict-match=0.9166

These full-set results are in the expected range for Qwen3-32B GSM8K accuracy and confirm the minimal Qwen3-32B support changes do not introduce an accuracy regression.

Test Plan

• Launched Qwen3-32B-FP8 with SGLang+ATOM, --page-size 16, and Qwen3 reasoning parser.
• Verified the service starts successfully without the previous RoPE/CUDA graph crash.
• Ran GSM8K A/B evaluation to isolate the low-score cause to generation length.
• Ran full GSM8K 3-shot evaluation with max_gen_toks=2048,max_length=8192.
• Confirmed SGLang service cleanup and GPU memory release after validation.