Accelerating LLM Inference with Parallel Draft Models (PARD)