[Performance] Speed up subgraph tracing for large models

vllm-project/llm-compressor · ★3,576 · issue #2981

This counts on your résumé — a merged PR here is verifiable proof of work.