feat: add GAIA benchmark for multi-step reasoning

[zeroasterisk]

Summary

• Adds GAIA (General AI Assistants) benchmark adapter following the existing Exgentic benchmark pattern
• Loads tasks from HuggingFace datasets (gaia-benchmark/GAIA, validation split)
• Supports 4 subset configs: 2023_all, 2023_level1, 2023_level2, 2023_level3
• Scores by exact match with standard normalization (lowercase, strip articles, remove punctuation, collapse whitespace)
• Reports per-level accuracy metrics in aggregation results
• Registers the benchmark in interfaces/registry.py with slug gaia

Structure

src/exgentic/benchmarks/gaia/
  __init__.py
  gaia_benchmark.py    # GAIASession, GAIAEvaluator, GAIABenchmark
  requirements.txt     # datasets>=2.0.0

Pattern mapping

Exgentic concept	GAIA implementation
Benchmark config	GAIABenchmark — subset selection, runner config
Session	GAIASession — loads one question, presents to agent, collects answer
Evaluator	GAIAEvaluator — lists task IDs, aggregates per-level accuracy
Finish action	submit with answer: str field
Scoring	Exact match after normalization (case-insensitive, whitespace/punctuation/article stripped)

Notes

• The GAIA dataset is gated on HuggingFace — users need an accepted access request and HF_TOKEN set
• Works with any Exgentic agent (tool_calling, claude_code, openai_solo, etc.)
• No additional tools are provided by default; agents use their own capabilities to reason through multi-step questions

Test plan

• Registry loads gaia benchmark correctly (validated via load_benchmark)
• All existing benchmarks still load after registry changes
• Answer normalization and exact-match scoring verified with unit assertions
• Evaluator returns correct task counts for all 4 subsets
• End-to-end run with a real agent (requires HF_TOKEN with GAIA access)

[zeroasterisk]

Just checking in — this is ready for review whenever you have bandwidth. Happy to address any feedback.