Fix MMMU VLM eval max_tokens for CoT prompt

[yhyang201]

Motivation

PR #21841 changed the MMMU eval prompt to use Chain-of-Thought instructions ("Think step by step before answering") but kept max_tokens=30. This is too short for models to complete their reasoning before outputting Answer: X, causing all VLM models to fail the nightly MMMU accuracy test since April 2.

Modifications

Increase max_tokens from 30 to 1024 in test/registered/eval/test_vlms_mmmu_eval.py.

Verification

Tested on H200 with all models in the nightly VLM MMMU eval:

Model	Score (max_tokens=30)	Score (max_tokens=1024)	Threshold	Status
Qwen2.5-VL-7B	0.28	0.56	0.33	PASS
MiMo-VL-7B-RL	0.30	0.49	0.28	PASS
deepseek-vl2-small	0.28	0.48	0.32	PASS
Qwen3-VL-30B-A3B	0.27	0.47	0.29	PASS
GLM-4.1V-9B-Thinking	0.29	0.47	0.28	PASS
GLM-4.5V-FP8	0.29	0.47	0.26	PASS
Qwen2-VL-7B	0.29	0.35	0.31	PASS
Janus-Pro-7B	0.30	0.33	0.285	PASS
InternVL2_5-2B	0.29	0.30	0.30	PASS

All models that can launch successfully now pass the accuracy threshold.

---

CI States

Latest PR Test (Base): ❌ Run #26991456983
Latest PR Test (Extra): ❌ Run #26991456867

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!