fix: constexpr HD for template param

tests/unit/test_qk_mma.cu

@@ -29,7 +29,7 @@ float cosine_sim(const float* a, const float* b, int n) {
 int main() {
     printf("=== tcgen05.mma QK GEMM Verification ===\n\n");
 
-    int HD = 64;
+    constexpr int HD = 64;
     int s_k = 128;
     int B = 1, H = 1;
     float scale = 1.0f / sqrtf((float)HD);