Reinforcement Learning API adds stochastic neural networks for hierarchical RL

AI Impact Summary

This change enables stochastic neural networks to be used within hierarchical reinforcement learning workflows, introducing stochastic latent variables that influence sub-policy selection across levels. It can improve exploration efficiency and policy quality, but also increases training variance and reliance on seeding, entropy tracking, and robust monitoring of non-deterministic outputs. Teams should extend their evaluation and deployment pipelines to compare stochastic versus deterministic baselines, adjust training workflows to support sampling strategies, and ensure model registries and monitoring dashboards capture entropy and policy performance for reproducibility.