NexaQuant: Llama.cpp-Compatible Model Compression with 100%+ Accuracy Recovery