Which large language model (LLM) parameter controls the number of possible next words or tokens considered at each step of the text generation process?