add gemm spec in kernel name

include/ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle.hpp

@@ -996,7 +996,8 @@ struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
             << MPerBlock << ", "
             << Gemm1NPerBlock << ", "
             << Gemm1KPerBlock << ", "
-            << B1K1 << ">";
+            << B1K1 << ", "
+            << getGemmSpecializationString(GemmSpec) << ">";
         // clang-format on
 
         return str.str();

include/ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm_xdl_cshuffle.hpp

@@ -913,7 +913,8 @@ struct DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle
             << MPerBlock << ", "
             << Gemm1NPerBlock << ", "
             << Gemm1KPerBlock << ", "
-            << B1K1 << ">";
+            << B1K1 << ", "
+            << getGemmSpecializationString(GemmSpec) << ">";
         // clang-format on
 
         return str.str();