Add MathArena 2025 final-answer benchmark suite

[taivu1998]

Summary

Adds five MathArena 2025 final-answer benchmark adapters with concise Evalchemy task names:

• Apex2025 for MathArena/apex_2025
• ApexShortlist for MathArena/apex-shortlist
• CMIMC2025 for MathArena/cmimc_2025
• BRUMO2025 for MathArena/brumo_2025
• SMT2025 for MathArena/smt_2025

The adapters share a single MathArenaFinalAnswerBenchmark implementation that mirrors the existing HMMT MathArena protocol: repeated deterministic generations, \boxed{} final-answer prompting, the vendored MathArena parser/checker, HMMT-style aggregate metrics, and rank-0-only result materialization.

What changed

• Added a shared eval/chat_benchmarks/matharena_final_answer_common.py helper for MathArena final-answer datasets.
• Added thin task wrappers for Apex 2025, Apex Shortlist, CMIMC 2025, BRUMO 2025, and SMT 2025.
• Added single-task configs with explicit max_tokens: 32768 and batch_size: 1.
• Added README coverage for the new MathArena 2025 final-answer suite.
• Added contract tests for wrapper shape, config parsing, task discovery, dataset normalization, row-count guards, metadata, pass@k guardrails, repeated generation, answer extraction, and repeated-run scoring.

Design notes

• The adapters intentionally share the common final-answer implementation instead of duplicating five near-identical benchmark classes.
• Concrete Evalchemy task/class names omit the MathArena prefix; dataset IDs and docs still name MathArena as the source.
• Native pass@k is guarded for these tasks for now because this implementation follows the existing HMMT repeated-run protocol rather than the newer native pass@k path.
• The shared loader validates expected row counts outside debug mode so upstream dataset drift fails loudly.
• Source-specific metadata is preserved where available: source for Apex-style datasets and problem_type for CMIMC, BRUMO, and SMT.

Validation

• python -m pytest -q -rs tests/test_matharena_final_answer_contract.py
  • 10 passed, 12 skipped
  • skips are from the local environment missing lm_eval
• python -m pytest -q tests/test_passk.py tests/resume
  • 103 passed, 6 skipped
• python -m py_compile tests/test_matharena_final_answer_contract.py
• line-length check for tests/test_matharena_final_answer_contract.py
• git diff --cached --check

Notes

The local shell prints a pre-existing virtualenvwrapper.hook_loader warning on command startup; commands above completed based on their exit status.