Distributed Training Process Isolation And Run Context Management

Scalable experiment tracking and model registry API.

Unique: Context manager pattern (`with Run()`) provides automatic run lifecycle management and process isolation without requiring explicit process coordination or queue management. Each process gets isolated run context, preventing metric collisions in distributed training.

vs others: Simpler than MLflow (no explicit run creation/closing) and more distributed-training-friendly than Weights & Biases (designed for multi-process logging without explicit process coordination)