Add custom sample prompts, more debug metrics, correctly reduce metrics

[shahbuland]

• Configs now have room for custom sample prompts rather than just randomly sampling from pipeline. You can now track how the models generations for these fixed prompts changes
• More debug metrics like KL divergence between model before/after a single RL training step and fraction of updates that were clipped by PPO clip on policy ratio
• Loss was not being reduced across GPUs before. Now it is.