Sheared LLaMA: Accelerating Language Model Pre-training via Structured Pruning