fix: add get_padding_offset to GPU ops stub

layers/utils.py imports get_padding_offset at module level when
CUDA is available. This caused collection error in CI.

Author: cloudforge1

tests/layers/test_fused_moe_deepgemm_backend.py

@@ -36,6 +36,7 @@
     "MoeWna16MarlinGemmApi",
     "tritonmoe_preprocess_func",
     "gptq_marlin_repack",
+    "get_padding_offset",
 ):
     if not hasattr(_gpu, _a):
         setattr(_gpu, _a, None)