vLLM TorchTPU: Zero-copy KV cache offloading and disaggregated serving