Add optimized CUDA kernels (RMSNorm, MLP) and patched Qwen3 modeling file

initial commit