refactor: remove 14 hardcoded benchmark branches from query/agent.py (CODE_HEALTH #2)

[duanyiqun]

CODE_HEALTH #2 (see docs/CODE_HEALTH.md in #37).

src/cognifold/query/agent.py:789-808 hardcodes a dispatch dict of 14 benchmark names mapping to 14 private _query_<bench>_qa methods (lines 811–1261) — several differing only by a magic max_nodes (40/15/5/30). Six entries (msc, qmsum, rgb, socialiqa, safetybench, futurex) refer to benchmarks whose directories no longer exist. query() itself is 263 lines; MemoryQueryAgent has 44 methods.

Cost: the shipped library carries leaderboard tuning; adding a benchmark means editing the core query engine; library users get 14 dead branches.

Fix direction:

• Introduce a QueryProfile dataclass (max_nodes, boosts, prompt template) accepted by query() / query_for_qa()
• Move per-benchmark values into benchmarks/*/ or configs/<name>_profile.yaml
• Delete the dispatch dict + 14 methods
• Do this behind tests (#CODE_HEALTH-1) and verify with --limit 1 runs of locomo/musique/tomi