Custom CUDA code runs Qwen3 30B A3B at 50 tok/s on RTX 5060 Ti

Custom CUDA and C++ code achieves 50-54 tok/s on an RTX 5060 Ti with 16GB VRAM, a 50% improvement over llama.cpp (33-34 tok/s). The optimization combines SOTA techniques to maximize memory bandwidth utilization.