PipeFill: Using GPUs During Bubbles in Pipeline-parallel LLM Training