End-to-End On-Device Quantization-Aware Training for LLMs at Inference Cost