TorchAO quantized Qwen3 models from PyTorch team, runnable in A100, H100 through vLLM and in mobile devices through ExecuTorch