Walk me through how attention scales with context length and why it matters for OpenAI-style long-context workloads.
