L2CEval: Evaluating Language-to-Code Generation Capabilities of Large Language Models