Exploiting Sparsity for Long Context Inference: Million Token on Commodity GPUs