MetaMask: Revisiting Dimensional Confounder for Self-Supervised Learning