VATLM: Visual-Audio-Text Pre-Training with Unified Masked Prediction for Speech Representation Learning