feat: Phase 4c — CNN + Ensemble architecture (multi-signal trading)

## Nouveaux modules

### src/ml/cnn/
- candlestick_encoder.py : CandlestickEncoder, fenêtres OHLCV z-score (N, 64, 5)
- cnn_model.py : TradingCNN — 3 blocs Conv1D(5→32→64→128) + BN + ReLU + GlobalAvgPool
- cnn_strategy_model.py : CNNStrategyModel, API identique à MLStrategyModel (train/predict/save/load)

### src/ml/ensemble/
- ensemble_model.py : EnsembleModel, poids {xgboost:0.40, cnn:0.60}, accord requis entre modèles

### src/strategies/cnn_driven/
- cnn_strategy.py : CNNDrivenStrategy(BaseStrategy), SL/TP ATR-based, fallback CNN_AVAILABLE=False

### src/strategies/ensemble/
- ensemble_strategy.py : EnsembleStrategy(BaseStrategy), auto-load XGBoost + CNN au démarrage

## Modifications

- trading.py : routes POST /train-cnn, GET /train-cnn/{job_id}, GET /cnn-models,
  POST /ensemble/configure, GET /ensemble/status + fix bugs (logging, _get_data_service, period_map)
- strategy_engine.py : support 'ml_driven' dans load_strategy()
- docker/requirements/api.txt : ajout torch>=2.0.0 + dépendances ML (scikit-learn, xgboost, lightgbm)

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

docker/requirements/api.txt

@@ -20,3 +20,12 @@ prometheus-client==0.19.0
 
 # Notifications
 python-telegram-bot==20.7
+
+# ML — requis pour MLDrivenStrategy (entraînement et prédiction dans l'API)
+scikit-learn==1.3.2
+xgboost==2.0.3
+lightgbm==4.1.0
+joblib>=1.3.0
+
+# ML — Deep Learning (CNN pour patterns chandeliers)
+torch>=2.0.0