ThanatosShinji

Support a new LLM model architecture in onnx-tool. Use when: adding a new transformer model (Qwen, DeepSeek, LLaMA, GPT, Phi, etc.) for graph building, profiling (MACs/params/memory), KV-cache analysis, or device latency projection. Covers the full workflow from HF config.json to ONNX export and performance profiling.