Refactor PPO objective function

[jon-tow]

This PR decouples the PPO loss function from the AcceleratePPOModel class for general reuse following @cat-state 's work in #75 . The following updates are also provided:

• Adds the unbiased KL-div estimates following http://joschu.net/blog/kl-approx.html
  • Related Issue: PPO Implementation Details - Checklist #53
• Removes repeated initialization of self.kl_ctl in accelerate_ppo_models constructor.
• Removes utils.modeling.clip_by_value as this functionality was added to PyTorch nearly 2 years ago.
• Introduces more PPO statistics to the logger for greater debugging information.

Test Run: https://wandb.ai/jon-tow/ppo-test

[LouisCastricato]

This looks great on my side, @Dahoas im open to merging this tonight if this is ok for you too?

[cat-state]

LGTM!