Add new benchmark suites

[Qubitium]

What changed

• Added runnable benchmark implementations for hle, supergpqa, hmmt_feb25, hmmt_nov25, hmmt_feb26, imoanswerbench, and livecodebench_v6.
• Registered capability-gated placeholders for swe_bench_verified, swe_bench_multilingual, swe_bench_pro, terminal_bench_2, claw_eval_avg, claw_eval_pass3, skillsbench_avg5, qwenclawbench,
  nl2repo, qwenwebbench, tau3_bench, vita_bench, deepplanning, tool_decathlon, mcpmark, mcp_atlas, and widesearch, with clear runtime-capability errors instead of misleading partial implementations.
• Exported the new suites through evalution.benchmarks and added integration metadata/baselines in tests/models_support.py.
• Added unit coverage plus standalone Llama 3.2 1B Instruct regression tests for the new runnable suites.
• Hardened math answer extraction to handle boxed answers, explicit final-answer lines, and inline math spans more reliably, using compiled pcre patterns.
• Added an optional apply_chat_template mode for HLE while keeping the default benchmark-faithful prompt path unchanged.

Validation

• PYTHON_GIL=0 pytest tests/test_scorers.py tests/test_hle.py tests/test_hmmt.py tests/test_imoanswerbench.py tests/test_livecodebench.py
• PYTHON_GIL=0 pytest tests/test_regex_backend.py
• CUDA_DEVICE_ORDER=PCI_BUS_ID CUDA_VISIBLE_DEVICES=4 PYTHON_GIL=0 pytest tests/models/test_hle_llama3_2_transformers.py tests/models/test_imoanswerbench_llama3_2_transformers.py tests/models/ test_livecodebench_llama3_2_transformers.py tests/models/test_supergpqa_llama3_2_transformers.py

Notes

• The low Llama 3.2 1B baseline scores on several newly added suites were checked directly; the main issue was task difficulty, not a broad scoring failure.
• HLE showed mild chat-template sensitivity, which is why the optional chat-wrapping mode was added for controlled A/B use.