Batch-ppo

Features:

- Split episodes into chunks for training. This reduces memory requirements when training from pixels and in some cases increases data efficiency.
- Use lambda variable initializers everywhere to support embedding the simulation into a larger graph.
- Upgrade to newest Gym version, including new environment names and dtypes for spaces.
- Support regularization losses returned by the network.

Improvements:

- Remove MuJoCo dependency from tests.
- Speed up smoke tests for faster iteration times.
- Enable continuous integration.

Bugs:

- Fix off-by-one bug in `FrameHistory` environment wrapper.

Features:

- Represent policies as tf.distribution objects, so that the algorithms are independent of the action distribution.

Improvements:

- Move reusable components into `agents.parts` package.
- Add nesting tools to handle nested tuples, lists, and dicts.

Bugs:

- Fix PPO not learning on GPU by placing the optimizer on the GPU.

Features:

- Use single optimizer for PPO to train shared feature layers better.
- Allow calling methods of the process environment.

Improvements:

- Improve default and MuJoCo configs.
- Report both training and evaluation scores.

Bugs:

- Likelihood calculation halved gradients for the action standard deviation.

Features:

- Policy networks are now defined as functions mapping sequences of observations to sequences of actions. As a result, feed forward policies are faster now, and memory based agents are easier to implement. Previously, networks were restricted to be defined as `RNNCell`s.
- All functions of the agent interface receive a tensor of agent indices now. This adds the flexibility to process observations in smaller batches. Previously, `perform()` and `experience()` was defined on data from all the environments.

Initial release.