Vesper: A Compact and Effective Pretrained Model for Speech Emotion Recognition