Rlgraph

- Added some new Components (MultiInputStreamNN, VariationalAutoEncoder, MultiLSTMLayer, JointCumulativeDistribution,
different supervised LossFunctions, Models, etc..), incl. test cases.
- Added container Space support to all Agents (python buffers had to be expanded).
- Bug fix in BernoulliDistributionAdapter and Policy (with bool actions): Was returning wrong
parameters (must be probs, not raw NN output).
- tf backend: Allow returning None now in graph_fn as valid DataOps.
- Bug fix in LSTM tf backend: LSTMLayer would not compile on tf versions < 1.13.
- Added possibility to flatten/split a graph_fn call "alongside" some given "(self.)property". This makes it easier to
split input args only to a certain extend (see e.g. distribution parameters in policy's API methods).

- Added a prototype for a debug visualization util that automatically
builds sub-sections of the (meta-)graph and visualizes this sub-graph
in the browser (as pdf) using GraphViz.
See FAQs for details on how to activate this feature.
Installing the GraphViz engine (and pypi `graphviz`) is not a requirement.
In the visualized sub-graph, where only the fault-relevant parts of
the Agent are shown to reduce information overload, one can see
immediately where Space (shape/type) problems occurred.
- Cleaner handling of exposing child API methods when calling:
`Component.add_components()`. The auto-generated (exposed) parent API
now has the same name and signature as the child's one (which makes it
more consistent when tracking incoming Spaces into API-input-args
(a Component's input-completeness is affected by this)).

- Fixed remaining issues with Keras-style NN assembly. This is now the recommended method for complex/multi-stream `NeuralNetworks`.
- Upgraded docker container to ubuntu:cosmic and tf1.13.1.
- Fixed bug in LSTMLayer occurring with newer tf version.

- Tuned learning tests for SAC and PPO.
- Fixed bug in Worker's reporting: `mean_episode_reward` was calculated
incorrectly as the return of the last episode (regardless of whether
this episode was completed or not).
- Added `mean_episode_reward_last_10_episodes` to Worker-returned reporting.

Fixed bug in PPOLossFunction affecting action spaces with shapes like (x, y, z, >1) and container action spaces.