Add Wang 2002 example

[mstimberg]

This is a reasonably fast implementation of Wang (2002), Probabilistic Decision Making by Slow Reverberation in Cortical Circuits, which many people are interested in (see e.g. this discussion on the forum. This example is particularly interesting, because the most straightforward implementation of the NMDA synapses – ignoring the special all-to-all connectivity structure with identical weights within subpopulations – will be very slow.

Actually, optimizing this model (which was mostly written by @wklausito) made me realize that for a model like this, C++ standalone now spends most of its time in creating random numbers (#1483). In C++ standalone, we are still using an old Mersenne-Twister implementation, whereas numpy has switched to more efficient RNGs by default.