API simplifications and better normalisation

• All functions creating traces over time with the bandpass filters have been removed from the class FeedforwardClosedloopLearning.
• There is now a new class FeedforwardClosedloopLearningWithFilterbank which has filter banks at every input.
• There is no longer a special input layer. The input layer contains activations and errors as any other layer.
• There is no longer a special "output layer" but simply the final layer is the output layer.
• The line follower demo is now part of this git so that you can see how the robot improves.
• loads of smaller fixes