Add pole balancing/reinforcement learning tutorial

[AlexisWis]

Added polebalancing tutorial which currently includes:
-Simulation Renderer
-Physics Simulation
-Classical Q-based agent

[clinssen]

See https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1003024:

We would like to point out the similarity of the TD-LTP learning rule to a reward-modulated spike-timing-dependent plasticity rule we call R-STDP [6], [16], [30]–[32]. In R-STDP, the effects of classic STDP [33]–[36] are stored into an exponentially decaying, medium term (time constant ), synapse-specific memory, called an eligibility trace. This trace is only imprinted into the actual synaptic weights when a global, neuromodulatory success signal is sent to the synapses. In R-STDP, the neuromodulatory signal is the reward minus a baseline, i.e., . It was shown [32] that for R-STDP to maximize reward, the baseline must precisely match the mean (or expected) reward. In this sense, is a reward prediction error signal; a system to compute this signal is needed. Since the TD error is also a reward prediction error signal, it seems natural to use instead of . This turns the reward-modulated learning rule R-STDP into a TD error-modulated TD-STDP rule (Figure 2A, bottom). In this form, TD-STDP is very similar to TD-LTP. The major difference between the two is the influence of post-before-pre spike pairings on the learning rule: while these are ignored in TD-LTP, they cause a negative contribution to the coincidence detection in TD-STDP.