feat(bench): implement GAIA dataset loader

[bug-ops]

Description

Implement the GAIA (HuggingFace leaderboard) dataset loader supporting levels 1, 2, and 3.

Part of epic #2827. See spec: .local/specs/zeph-bench/spec.md section 6.

Scope

• GaiaLoader parsing GAIA JSON schema; support --level 1|2|3 filter flag
• Accuracy evaluator (exact match normalized per GAIA spec)
• Download from HuggingFace Hub; cache to ~/.local/share/zeph/bench/gaia/
• Unit tests with a synthetic fixture

Acceptance Criteria

• Loader parses GAIA schema for all three levels
• --level filter restricts scenarios correctly
• Accuracy evaluator matches GAIA normalization rules on fixture
• zeph bench download --dataset gaia works