Benchmarking Multimodal LLMs in Real-World Ultra-High-Resolution Remote Sensing with Human-Verified Annotations