Implement Phase 8: Offline AI Model Training & Evaluation

[takurot]

Summary

This PR implements the offline training and evaluation pipeline for the AI Sidecar (Phase 8), enabling the system to learn optimal caching policies from query logs.

Changes

• P8-1: Training Pipeline: Added to train a GBDT model (XGBoost/Sklearn) on system metrics (, , 'latency' must be run as root..., ) and export it to ONNX.
• P8-2: Evaluation Pipeline: Added to simulate the model's impact on historical data, estimating Cost Savings and P99 improvements.
• P8-3: ONNX Validation: Added robust ONNX structure and runtime inference checks to ensuring model validity before deployment.
• Dependencies: Added , , , to .

Verification

• Training: Verified runs successfully on synthetic/logged data.
• Evaluation: Verified produces a simulation report (Estimated ~27% P99 improvement on test logs).
• Tests: Regressions tests passed ().