Allow better scaling of the deposition with number of threads

[RemiLehe]

This pull request introduces two changes, in order to have better scaling of the deposition routines with respect to the number of threads:

• The dimension of the global arrays (which duplicate data for each thread) are swapped (so that the thread index is the slowest index)
• The allocation of arrays that are local to each thread is removed

It turns out that on a Haswell 32-core CPU, this improves the scaling with threads quite dramatically.