Fix Usage example: use apply_chat_template for instruction-tuned model

Add optimized CUDA kernels (RMSNorm, MLP) and patched Gemma4 modeling files

initial commit