Quest: Query-Aware Sparsity for Efficient Long-Context LLM Inference