Context Rot: How increasing input tokens impacts LLM performance