Does quantization affect models' performance on long-context tasks?