From Bits to Rounds: Parallel Decoding with Exploration for Diffusion Language Models