Towards Free Normalization: Fusing Normalization into GEMM and Attention Kernels