fix: syntax error — move triton imports before docstring, remove orphan @triton.jit

patches/deepseek_v4.py

@@ -244,16 +244,16 @@ class DeepseekV4FP8Config(Fp8Config):
         return isinstance(layer, FusedMoE) and self.expert_dtype == "fp4"
 
 
-@triton.jit
+import triton
+import triton.language as tl
+import torch
+
 """
 NVFP4 staging kernel — full FP4 (E2M1) activations + UE4M3 block16 scales.
 
 The mxf4nvf4 PTX instruction requires BOTH A and B to be FP4 (E2M1 packed).
 This kernel quantizes BF16 activations → E2M1 packed uint8 with UE4M3 scales.
 """
-import triton
-import triton.language as tl
-import torch
 
 
 @triton.jit