Keras-rl

- Fix Tensorboard support
- Improve documentation

- Add support for current version of Keras and Gym

- Removes legacy code and now requires `Keras >= 2.0.7` (https://github.com/matthiasplappert/keras-rl/pull/157)
- Adds two new policies (https://github.com/matthiasplappert/keras-rl/pull/156, https://github.com/matthiasplappert/keras-rl/pull/122)
- Fixes a few problems with replay memory sampling being incorrect for certain edge cases (https://github.com/matthiasplappert/keras-rl/pull/138)
- Fixes a problem with dropout in actor model for DDPG (https://github.com/matthiasplappert/keras-rl/pull/150)

- Require Keras < 2.0.7 due to non-trivial and breaking API change in optimizer code

- Full compatibility with the recently released Keras 2
- New `SarsaAgent` implementation
- New `train_policy` on `DQNAgent` and `SarsaAgent` to allow for different policies at training and test time
- Minor clean-ups

- Fix version mix-up on release, otherwise identical to version `v0.2.1`