Fix CUDA stream: use c10::cuda::getCurrentCUDAStream() directly in kernel launch

dsv4/kernels/cuda/blackwell_swizzle.cu

@@ -104,12 +104,12 @@ void launch_blackwell_swizzle(
 // Pybind11 bindings for torch.utils.cpp_extension.load
 PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
     m.def("blackwell_swizzle_32_4_4", [](at::Tensor input, at::Tensor output, int32_t rows, int32_t cols) {
-        auto stream = at::cuda::getCurrentCUDAStream().stream();
-        launch_blackwell_swizzle(
+        auto stream = c10::cuda::getCurrentCUDAStream();
+        blackwell_swizzle_32_4_4_kernel<<<
+            (rows * cols + 255) / 256, 256, 0, stream>>>(
             input.data_ptr<uint8_t>(),
             output.data_ptr<uint8_t>(),
-            rows, cols,
-            stream
+            rows, cols
         );
     }, "Blackwell 32_4_4 scale swizzle");
 }