fix: restore cuda_bf16.h include

dsv4/kernels/attention/fmha_sm100.cuh

@@ -30,14 +30,8 @@
 #pragma once
 
 #include <cuda_runtime.h>
-#include <cstdint>
-
-// __nv_bfloat16 is a built-in type on CUDA 13+
-// cuda_bf16.h may have C++17 compatibility issues with some CUDA versions;
-// only include it in device code where it's guaranteed to work.
-#if defined(__CUDA_ARCH__)
 #include <cuda_bf16.h>
-#endif
+#include <cstdint>
 
 // CUTLASS C++ includes (CUDA device code only)
 #if defined(__CUDA_ARCH__)