fix: federated learning memory leaks and stability

[sigfawn]

This PR addresses stability issues in the Federated Learning service.

1. Memory Leak Prevention

• Problem: EphemeralLearningAgent accumulated task history indefinitely until aggregation occurred. If the coordinator was offline or the interval failed, this caused unbounded memory growth.
• Fix: Added a maxHistorySize (default 1000) cap. The history queue now automatically drops the oldest tasks when full.

2. Process Lifecycle Management

• Problem: FederatedLearningManager started a setInterval that kept the Node.js event loop alive indefinitely, causing CLI commands and tests to hang instead of exiting.
• Fix: Applied .unref() to the aggregation timer, allowing the process to exit naturally when no other work is pending.

Result: Safer, leak-proof distributed learning agents.