Verifying LLM Inference to Detect Model Weight Exfiltration