Add dump_metric to MMMU, lm-eval, and NeMo Skills eval paths

[hnyls2002]

Summary

• Add dump_metric calls to three eval paths that were missing them: MMMU (lmms-eval), lm-eval harness, and NeMo Skills (mmmu-pro)
• All eval paths through run_eval.py already had dump_metric; this covers the remaining ones
• dump_metric is silent on failure — no risk to existing tests

This is Phase 2 of the eval unification plan started in #21667 (Phase 1: GSM8K unification). The goal is to ensure all eval paths emit dump_metric output, laying the groundwork for regression detection infrastructure in future phases.

Changes

File	What
kits/mmmu_vlm_kit.py	MMMUMixin.test_mmmu + MMMUMultiModelTestBase._run_vlm_mmmu_test
kits/lm_eval_kit.py	LMEvalMixin.test_lm_eval per task/metric
accuracy_test_runner.py	_run_nemo_skills_eval after score parsing

Test plan

• CI passes (no functional change — dump_metric is additive and silent-fail)

[gemini-code-assist]

Code Review

This pull request integrates metric dumping into several test kits, including the accuracy test runner, LM evaluation kit, and MMMU VLM kit, to track evaluation scores. A review comment suggests standardizing the labeling schema in the LM evaluation kit to match the other modules by using the task name for the 'eval' label and adding an 'api' label for the framework.

[gemini-code-assist]

The labeling schema here is inconsistent with the other evaluation paths modified in this PR. In mmmu_vlm_kit.py and accuracy_test_runner.py, the eval label represents the benchmark/dataset name (e.g., "mmmu" or "mmmu-pro") and the api label represents the framework/runner (e.g., "lmms-eval" or "nemo-skills").

In this file, eval is set to "lm-eval" and the benchmark is stored in a separate task label. To maintain consistency across the metrics collected from different kits, consider using the task name for the eval label and adding an api label set to "lm-eval".

Suggested change

	labels={
	"model": eval_config.get("model_name", ""),
	"eval": "lm-eval",
	"task": task["name"],
	},
	labels={
	"model": eval_config.get("model_name", ""),
	"eval": task["name"],
	"api": "lm-eval",
	},

[hnyls2002]

/tag-and-rerun-ci