feat: add ARC-AGI-1, ARC-AGI-2, and BrowseComp benchmark tasks

[Luodian]

Summary

• Add ARC-AGI-1 image benchmark task using community dataset mertaylin/arc-agi-images (400 eval samples, exact grid match)
• Add ARC-AGI-2 image benchmark task using community dataset vincentkoc/arc-agi-2 (120 eval samples, interleaved demo images)
• Add BrowseComp text benchmark task from OpenAI (1,266 encrypted samples, LLM-as-judge with exact match fallback)

Details

ARC-AGI-1 (lmms_eval/tasks/arc_agi_1/)

• Dataset: mertaylin/arc-agi-images (eval split)
• Uses stacked demo composite image + test input image
• Metric: arc_agi_1_acc (exact grid match, JSON array-of-arrays parsing)

ARC-AGI-2 (lmms_eval/tasks/arc_agi_2/)

• Dataset: vincentkoc/arc-agi-2 (evaluation split)
• Interleaves train input/output images as few-shot demos
• Metric: arc_agi_2_acc (exact grid match with _normalize_grid type coercion)

BrowseComp (lmms_eval/tasks/browsecomp/)

• Dataset: OpenAI Azure blob CSV (XOR-encrypted to prevent contamination)
• Decryption at eval time using SHA256-derived key
• LLM-as-Judge scoring via lmms_eval/llm_judge when API configured; falls back to exact match
• Custom aggregation with per-topic accuracy breakdown
• Metric: browsecomp_acc

Verification

• All 3 tasks discoverable via --tasks list ✅
• Python syntax check passed ✅
• pre-commit run (black + isort) passed ✅

Linear Issues

Resolves LMM-306, LMM-307, LMM-273