Understanding Emergent Abilities of Language Models from the Loss Perspective