Methods should (maybe) not return Observation* and Action*

[sofian]

The reason being that in no case are these pointers generated on the spot. They are thus confusing. This may also result in memory consuming implementations, for instance when one has to "save" those pointers for later use.

Instead, Agent should just have an Action* and Environment an Observation*.

[sofian]

There is an issue though, which is that by using a member class pointer, we need to dynamically allocate the Action or Observation object, which is a bit tricky on AVR. However, since we already do it for NeuralNetwork + we have a StaticAllocator mechanism, then, why not.