Add AMC24 and AMC25 text-ready benchmarks

[taivu1998]

Summary

Adds AMC24 and AMC25 as committed, local Evalchemy chat benchmarks:

• vendors reviewed AMC 12 A/B text-ready subsets for 2024 and 2025
• includes source-figure rows only when the prompt text fully specifies the problem without needing the diagram
• adds source manifests with canonical MAA/AoPS provenance, Hugging Face mirror cross-checks, row counts, manually included self-contained figure rows, and remaining exclusions
• adds a shared AMC benchmark helper for prompt rendering, repeated generation, boxed-answer extraction, and scoring
• adds lightweight AMC answer utilities so normalization/scoring can be tested without requiring the full lm_eval stack
• documents AMC24/AMC25 in the README benchmark list

The dataset-generation script and planning files are intentionally excluded from this PR; only the committed benchmark data, runtime adapters, shared helpers, tests, and README changes are included.

Data Scope

• AMC24 includes 48 text-ready rows.
• AMC25 includes 49 text-ready rows.
• Remaining excluded rows:
  • AMC24: 2024-amc12a-18, 2024-amc12a-22
  • AMC25: 2025-amc12a-05
• Rows with source figures are marked with has_figure=true plus figure_handling=source_figure_omitted_text_self_contained when the diagram is redundant for the converted text prompt.
• Runtime benchmark execution uses local committed JSONL only, with no network fetch.

Validation

• uvx --with pytest pytest tests/test_amc_benchmarks.py -q
• uvx ruff check eval/chat_benchmarks/amc_common.py eval/chat_benchmarks/amc_utils.py eval/chat_benchmarks/AMC24/eval_instruct.py eval/chat_benchmarks/AMC25/eval_instruct.py tests/test_amc_benchmarks.py
• uvx ruff format --check eval/chat_benchmarks/amc_common.py eval/chat_benchmarks/amc_utils.py eval/chat_benchmarks/AMC24/eval_instruct.py eval/chat_benchmarks/AMC25/eval_instruct.py tests/test_amc_benchmarks.py
• python3 -m py_compile eval/chat_benchmarks/amc_common.py eval/chat_benchmarks/amc_utils.py eval/chat_benchmarks/AMC24/eval_instruct.py eval/chat_benchmarks/AMC25/eval_instruct.py tests/test_amc_benchmarks.py
• structured data-contract check for row counts, exact exclusions, manual inclusion IDs, and figure-handling metadata

Note: uv run pytest tests/test_amc_benchmarks.py -q still fails before collection on the existing project metadata issue: fschat @ file:eval/chat_benchmarks/MTBench uses a relative file dependency that uv cannot resolve while building the package.