Learning is Forgetting: LLM Training As Lossy Compression