[None][test] Add InferenceMAX-style GSM8K accuracy eval mode

[zheyuf]

Summary

Adds an InferenceMAX-style GSM8K accuracy eval mode so model accuracy can be validated under the exact protocol published at inferencex.semianalysis.com/evaluation (SemiAnalysisAI/InferenceX), directly comparable to the vLLM/SGLang reference numbers there.

The existing in-repo gsm8k task measures several points below InferenceMAX's leaderboard numbers (MiniMax-M3-NVFP4: ~90.5 vs ~95.5) because it applies no chat template, runs no thinking, and caps output at 256 tokens.

The new mode is model-agnostic and easy to extend to other InferenceMAX-tracked models (DSR1, GPT-OSS, Kimi, GLM, ...). This is a port of the side-branch commit feat/m3_with_msa #16711 onto main.

What's in the PR

• tensorrt_llm/evaluate: GSM8KInferenceMax evaluator + lm_eval_tasks/gsm8k_inferencemax/gsm8k_inferencemax.yaml (mirrors InferenceX utils/evals/gsm8k.yaml: chat template, 5-shot multiturn, #### [number] format instruction, max_gen_toks: 12288, strict #### N extraction). LmEvalEvaluator gains shuffle_dataset / fewshot_random_seed so the task reproduces the lm-eval CLI's exemplar selection (unshuffled train split, sampler seed 1234); defaults for every other task are unchanged.
• trtllm-eval gsm8k_inferencemax CLI subcommand.
• Accuracy harness: GSM8KInferenceMax task in accuracy_core.py; reference rows may now carry an explicit threshold (replaces the hypothesis-test threshold with a fixed floor, reported as such) and a free-text reference_note (provenance of the reference score). Both are optional and no existing row uses them.
• TestMiniMaxM3::test_nvfp4 is parametrized by eval_mode: default = MMLU + GSM8K as before; inferencemax = chat-format GSM8K at 16k context with max_batch_size=64 (InferenceMAX evaluates a serving endpoint at client concurrency 64).
• Test lists: the QA list (llm_function_core.txt) runs the inferencemax variants (TIMEOUT 180, thinking output dominates in eager mode); pre-merge l0_dgx_b200 keeps eval_mode=default. Existing waives are renamed to the new ids.

Accuracy thresholds (references/gsm8k_inferencemax.yaml)

Reference scores are the published InferenceMAX vLLM B200 FP4 numbers (2026-07); a strict explicit floor is used to guard model quality during bring-up / perf work.

Test	Reference (source)	Threshold
test_nvfp4[use_msa=False-eval_mode=inferencemax]	95.46, vLLM B200 FP4 spec-off mean [94.84-95.91] n=13	94.0
test_nvfp4[use_msa=True-eval_mode=inferencemax]	95.46, same row, FP8 KV cache	94.0

Side-branch measurements (4x B200, TP4/EP4, 2026-07-22): 95.53 (triton, BF16 KV) and 95.68 (MSA, FP8 KV); run-to-run SD 0.26, per-run lm-eval stderr 0.58. InferenceMAX's own gate for this task is 90.0.

Differences from the side-branch commit

• The Eagle3 arm (test_nvfp4_eagle3[...-eval_mode=inferencemax]) and its reference row (published vLLM B200 FP4 Eagle3 mean 95.49) are not in this PR: test_nvfp4_eagle3 does not exist on main yet. They follow with [TRTLLM-14093][feat] Eagle3 support for MiniMax-M3 on the MSA backend #16021 (Eagle3 for MiniMax-M3 on the MSA backend), which will pick up eval_mode on top of this PR.

Test coverage

• accuracy/test_llm_api_pytorch.py::TestMiniMaxM3::test_nvfp4[use_msa=*-eval_mode=inferencemax] (QA list).
• accuracy/test_llm_api_pytorch.py::TestMiniMaxM3::test_nvfp4[use_msa=True-eval_mode=default] (pre-merge l0_dgx_b200, currently waived on main under nvbugs/6601633; waive renamed).

To reviewers

• As on the side branch, the QA list switches the MiniMax-M3 test_nvfp4 entries to InferenceMAX mode instead of adding the variants alongside the default ones (to keep QA time flat). Happy to keep both if preferred.
• The mode has only been measured on the side branch so far; CI (/bot run) on this PR is the first run on main.

PR Checklist

• PR title follows the [JIRA/NVBUG/None][type] format
• Commits are signed off (DCO)
• pre-commit clean (isort / yapf / test-list validators)
• CI (/bot run)