CUDA-L2: Surpassing cuBLAS Performance for Matrix Multiplication Through RL