Refactor transformation function

[pdeubel]

Solves #13. Unfortunately I also did some work on #47 which is already in this PR: #49. So this PR stays a draft until #49 is merged then I will resolve the conflicts.

TODO here:

• [x] Write some tests for the transformations
• [x] Wait until #49 is merged, then solve code conflicts

---

This PR does the following:

• Removed all the transformations of observations etc from brain classes
• Introduces parse_brain_input_output() method that takes either an input for the brain (i.e. observation from an environment) or an output from the brain and parses it into a suited format, depending on the target Space.
  • If the Space is of type spaces.Box: If a brain input is parsed nothing is done to it, if a brain output is parsed it gets reshaped to the dimensions of the output Space
  • If the Space is of type spaces.Discrete: If a brain input is parsed the input is encoded in a one-hot vector, if a brain output is parsed the argmax is taken from the vector to provide a single value to the environment
  • If the Space is of type spaces.Tuple: If a brain input is parsed, it will be iterated through the subspaces and for each subspace the parse_brain_input_output() is called recursively. The resulting parsed inputs then get flattened and concatenated so that the Brain will become a one-dimensional NumPy Array. If a brain output is parsed the output will be reshaped to match the Tuple output space, i.e. the subspaces of the Tuple have certain dimensions and a tuple is created where its entries will correspond to these dimensions.
• Removed normalization of brains. Removed the method for this as well as the Brain configuration parameters

[bjuergens]

The resulting parsed inputs then get flattened and concatenated so that the Brain will become a one-dimensional NumPy Array

this probably breaks CNN-brains. Currently my implementation of CNN-Brains works very badly anyways, so we may just drop them.