Court filings suggest that Meta staffers discussed using copyrighted materials to train Meta’s AI models, including models in its Llama family.