Diffusion-Inspired Masked Fine-Tuning for Knowledge Injection in Autoregressive LLMs