feat: Supersonic evaluator, Cactus Kev hash, repo split cleanup

Evaluator overhaul:
- Add Supersonic LUT trie evaluator (32M-entry, O(n) hand evaluation)
- Add Cactus Kev perfect-hash 5-card evaluator (prime-product hashing)
- 3-tier evaluation priority: Supersonic → Cactus Kev → Algorithmic
- DashMap-cached 7-card evaluation for lock-free concurrent access
- Rewrite JumpTable around CactusKevEvaluator + DashMap cache

Bot framework:
- Add StrategyBot adapter (Strategy trait → HoldemPlayer)
- Add register_bot_type() for plugin-style bot registration
- Add strategy module with modular composable strategies

Repository split:
- Remove holdem_bots from workspace (moved to private super_marvin repo)
- Remove holdem_bots dependency from holdem_sim
- Update cash_game.toml to use built-in bots only
- Clean up .gitignore (remove stale holdem_bots exclusion)
- Delete plans/ directory (internal planning artifacts)

Documentation:
- Add ARCHITECTURE_SUPERSONIC_EVALUATOR.md
- Add REFERENCES.md
- Update ROADMAP.md and VISION.md
- Remove outdated docs (ACHIEVEMENTS, RAKE, TOURNAMENT_RAKE, meerkat-alignment)

docs/VISION.md

@@ -14,7 +14,7 @@ Create a **comprehensive poker AI ecosystem** with two main components:
 
 - **Configurable bots** (Rule-based, Nash EQ, Statistical, Hand-reading, Reinforcement Learning)
 - **Mass parallel simulations** with GPU acceleration (e.g., for turn and river rollouts)
-- **Memory-efficient evaluation**: 79% memory reduction enables larger-scale simulations
+- **Multi-tier hand evaluation**: Supersonic LUT trie → Cactus Kev perfect-hash → Algorithmic fallback, with DashMap-cached 7-card evaluation
 - **TOML-driven configuration** for all parameters of bots and simulations
 - **Comprehensive result analysis** and bot evolution