[WIP] Fix default judge model selection conflict in run.py and tools.py

[TianhaoLiang2000]

Summary

This PR centralizes the default judge model selection logic into vlmeval/judge.py and makes both run.py and vlmeval/tools.py::EVAL() use the same source of truth.

Changes

• Added get_default_judge_model() in vlmeval/judge.py.
• Updated run.py::get_judge_kwargs() to reuse the shared default judge resolver.
• Updated vlmeval/tools.py::EVAL() to use the same resolver instead of its older hardcoded defaults.
• Removed the stale chatgpt-0125 default for MCQ / Y-N evaluation through tools.py::EVAL().
• Kept compatibility for both MMBench_Video and MMBench-Video dataset naming.

Testing

• python -m py_compile run.py vlmeval/tools.py vlmeval/judge.py
• pre-commit run passed during commit.
• pre-commit run --all-files was also run; it fails only on pre-existing flake8 issues outside this PR's changed files.

[mzr1996]

Please move the judge model selection to the benchmark class (for example, a class attribute), and this function will use the class attribute and dataset type to determine the final judge model name.