Demystifying ARM SME to Optimize General Matrix Multiplications