Self-prompting and cross-model consensus enable reproducible data extraction from scientific literature with large language models