Improve memory allocation of integration step

[enekomartinmartinez]

Currently, integration process works as follows.

1. Create a big numpy.array of all the stateful objects
2. Create a big numpy array of all ddt
3. Update the 1st numpy array of all objects using the 2nd

Instead of creating such a big numpy array, we should try computing the updating values inside each array in pre-allocated memory space. This should improve specially the performance of big models.