APB: Accelerating Distributed Long-Context Inference by Passing Compressed Context Blocks across GPUs