feat: expand datasets to 55 samples and add CI regression command

[rudrakshkarpe]

Summary

• 5 new benchmark datasets with 52 additional samples (55 total across all datasets):

  • encoding-attacks (10 samples): base64 encoded instructions, unicode tricks, HTML entity/comment abuse, URL encoding, zero-width chars, markdown image exfiltration
  • false-positives (10 samples): benign content that superficially resembles attacks — security training docs, legitimate CEO emails, webhook configs, audit reports, i18n content
  • tool-abuse (10 samples): targeted exploitation of each forbidden tool (send_external_webhook, delete_memory_note, mark_safe_to_ignore) plus multi-tool chains
  • authority-escalation (10 samples): CEO impersonation, fake system admin overrides, legal counsel demands, stacked multi-authority claims
  • data-exfiltration (12 samples): direct/disguised webhook sends, memory poisoning, markdown beacons, social engineering chains, cover-up sequences

• CI regression command (dual-llm-bench ci):

  • dual-llm-bench baseline <traces.jsonl> saves current scores as a baseline JSON file
  • dual-llm-bench ci <traces.jsonl> --baseline .dlb-baseline.json compares against baseline and exits non-zero if any metric regresses beyond a configurable threshold (default 5%)
  • Clear diff table output showing baseline vs current per metric

• Generalized dataset loader to support multiple datasets via a registry dict instead of hardcoded if check

• Widened BenchmarkSuite.from_builtin() to accept str instead of Literal["pycon-core"]

Closes #1, closes #9

Test plan

• All 6 datasets load and validate (test_dataset_loads_successfully parametrized)
• All sample IDs unique across the entire corpus
• Attack samples have forbidden_tools and attack_markers; benign samples have neither
• Total sample count >= 55
• baseline command creates valid JSON file
• ci command passes when scores match baseline
• ci command fails with exit code 1 on regression
• ci handles missing baseline file
• Custom threshold (--threshold 1.0) allows large regressions to pass
• Existing tests (metrics, CLI) continue to pass
• ruff check clean, mypy --strict clean