revert: 移除所有 torch.compile（四战全败），回到稳定版 58.49

torch.compile 全模式验证：
- reduce-overhead: 199s (+126%)
- default 全模型: 118s (+34%)
- default Expert: 108.6s (+23%)
- dynamic=True: 102.6s (+17%)
MoE 动态路由 + 可变序列长度，与任何 JIT 编译不兼容

代码/code/infer.py

@@ -504,14 +504,6 @@ def load_model(ckpt_path, device='cuda:0'):
         print(f"[WARNING] Checkpoint {ckpt_path} not found, using random weights")
 
     model.to(dev)
-
-    # === torch.compile + dynamic=True：告知编译器形状可变，避免重编译 ===
-    try:
-        model = torch.compile(model, dynamic=True)
-        print(f"[INFO] torch.compile applied (dynamic=True)")
-    except Exception as e:
-        print(f"[WARNING] torch.compile failed ({e}), using original model")
-
     model.eval()
     print(f"[INFO] Model ready. Device: {dev}")