Qwen3 30B A3B runs 50 tok/s on RTX 5060 Ti with custom CUDA code

Custom CUDA and C++ implementation achieves 50-54 tok/s for Qwen3-30B-A3B at float8 on a 16 GB VRAM RTX 5060 Ti, roughly 50% faster than llama.cpp (33-34 tok/s). Speedups come from combining SOTA kernel optimizations.