Improve late interaction/late chunking context window size once https://github.com/abetlen/llama-cpp-python/issues/1762 is fixed.