LoBaSS: Gauging Learnability in Supervised Fine-tuning Data