Finish submission checker log parsing in both endpoint logs and loadgen logs

[zihaok]

MR Enhancement Summary

• Enhanced endpoint submission parsing to read latency, TTFT, and TPOT from early_stopping_percentiles instead of the normal percentile fields. please see this MR in endpoint repo feat(metrics): opt-in MLPerf early-stopping percentile estimates (LoadGen parity) endpoints#417 that implement early stopping in endpoint
• Updated endpoint LLM latency validation so LLM models collected both in endpoint and loadgen use TTFT/TPOT limits, while non-LLM models collected using endpoint or loadgen continue to use end-to-end latency checks.
• Bugfix: Added gpt-oss-120b to the LLM model set so it follows the same TTFT/TPOT validation path as DeepSeek-R1.
• Fixed endpoint accuracy sample-count handling so accuracy coverage is derived from accuracy/accuracy_results.json, not from performance n_samples_issued or n_samples_completed.
• Added scenario-specific Qwen3-VL accuracy sample-count handling to avoid applying one dataset-size requirement across all endpoint scenarios. Server: 48k, Interactive: 8K
• Tightened endpoint compliance-test handling to only include test04 in Wan, instead of checking every benchmark's audit test folder. gptoss and dsr1 doesn't need separate audit test.
• Improved endpoint parser robustness for nested JSON paths and percentile keys such as 99.0 and 99.9.
• in Endpoint log parsing, enforce accuracy dataset has a sample count that is the same as dataset whole size

[github-actions]

MLCommons CLA bot All contributors have signed the MLCommons CLA ✍️ ✅

[pgmpablo157321]

I have tested it with the sample logs and looks good