Towards Latent Masked Image Modeling for Self-Supervised Visual Representation Learning