Generative Pre-training for Speech with Flow Matching