Beyond Correctness: Benchmarking Multi-dimensional Code Generation for Large Language Models