Making Self-supervised Learning Robust to Spurious Correlation via Learning-speed Aware Sampling