feat(bench): implement LongMemEval dataset loader and evaluator

[bug-ops]

Description

Implement the LongMemEval dataset loader, scenario iterator, and evaluator (exact match + F1).

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

Scope

• LongMemEvalLoader that reads the official dataset JSON format and produces Vec<Scenario>
• Correct handling of multi-session scenarios: when session_id changes within a scenario, trigger a memory reset mid-scenario (new session boundary)
• LongMemEvalEvaluator implementing exact-match and token-F1 metrics against ground-truth answers
• Dataset download: fetch from official source URL, cache to ~/.local/share/zeph/bench/longmemeval/
• Cache validation: SHA256 checksum of downloaded archive
• Unit tests with a small synthetic scenario fixture (3 turns, known ground truth)

Acceptance Criteria

• Loader correctly parses official LongMemEval JSON schema
• Multi-session boundary triggers memory reset
• F1 evaluator matches reference implementation on 3-example fixture
• Download and cache work end-to-end (zeph bench download --dataset longmemeval)
• Unit tests pass without network access (fixture-based)