Beyond Fully Random Masking: Attention-Guided Denoising and Optimization for Diffusion Language Models