Parallelized Planning-Acting for Efficient LLM-based Multi-Agent Systems