More Parallelization Efforts

[antoinecarme]

Following #115 , we can go further in parallelizing PyAF computations. One can :

1. Use more CPUs in training (improves cross validation training time)
2. Allow using multiple signals in the same engine (multiple signals sharing the same CPUs)
3. Use more CPUs in forecasting
4. Use one engine to train all nodes of a hierarchy (point 2).

[antoinecarme]

impact of the APIs :

One can now use a list iSignals as an argument for Engine.Train, while keeping a strict backward compatibility (a single iSignal is still supported).