Add optimized CUDA kernels (RMSNorm, MLP, Attention) and patched Llama-3.1 modeling file

initial commit