Beyond Autoregression: An Empirical Study of Diffusion Large Language Models for Code Generation