A Systematic Analysis of Large Language Models as Soft Reasoners: The Case of Syllogistic Inferences