Add Gemma 4 support

[jethac]

Summary

• Add a Gemma4 model adapter with transformers and vLLM inference paths
• Register Gemma 4 instruction-tuned model variants in the Gemma model group
• Normalize local images as RGB PIL images and strip thinking blocks from returned responses for evaluation

Tests

• python -m py_compile vlmeval/config.py vlmeval/vlm/gemma.py
• python -m pre_commit run --files vlmeval/config.py vlmeval/vlm/gemma.py vlmeval/vlm/init.py