FLOAT: Generative Motion Latent Flow Matching for Audio-driven Talking Portrait