Investigate asynchronous NN model update

[jardinetsouffleton]

The goal of this issue would be to investigate the possibility of de-coupling the model used to make desicions (value assignment) from the model that is being updated. As of now, the backpropagation occurs at every update_freq time steps (after update_freq decisions). When the model is being updated by the backpropagation, the trajectory generation must stop. However, as the trajectory generation is done one sample at a time, there is a lot of space remaining on the GPU to compute the backprop asynchronously. Doing this would allow us to generate more trajectories and possibly speed up the resolution of problems