Becca

Some bugfixes and minor improvements over 0.10.0 to make becca suitable for integration with, for example, OpenAI Gym.

Major changes:
* The Discretizer was added. Now sensor inputs don't need to be pre-discretized or scaled.
* A full visualization was added and broken out into the becca_viz package.

There are also minor changes throughout.

Significant improvements to version 9.0, courtesy SethHWeidman.

This merge is long overdue.

Becca 9 includes a re-engineered reinforcement learner. It is based the reinforcement learner from Becca 7, but with some major bugs worked out and more speed. I undid most of the changes from version 0.8.2. Becca 9 outperforms all previous versions on the `becca_test` worlds. There is still a long list of next problems to tackle, but it is worthy of a release.

- All inputs (not just nonbundle activities) are used to train new bundles.
- **Increases** in element activities are used to learn sequences, rather than absolute values.
- Instead of choosing one goal every time step, it is now permissible not to choose a goal. Sequences are chosen as goals probabilistically based on their expected value.

Additional context

I just hit a roadblock in the development of Becca 8. The approach I'm using estimates reward for state-action pairs (similar to a traditional value function). I realized I need to revert back to the approach in Becca 7 where I used state-action-state tuples as my basis. This allows for state predictions, world modeling and flexible multi-step planning. Becca 8 taught me several things about how to best represent and combine inputs. Now I need to move on to Becca 9 and integrate the best of both 7 and 8. I'll release the current state as Becca 0.8.2 and move on.

This release includes a ground-up reworking of all the algorithms. There are some fundamental changes.
- Every feature that is created is now a short temporal sequence. This is a way to build time into the fabric of Becca's perception.
- The deep learning and reinforcement learning are now fully integrated, rather than two modular blocks. The

There are some engineering changes that have been made too.
- The test worlds have also been split out into a separate package.
- Becca and the test worlds are now on pypi and installable using pip.

Enjoy!