What is an LLM context window, and how do you handle tasks that require more context than the model can take at once?