Implement `mean` reduction method

[pastewka]

This would allow simple parallelization of many of the scalar parameters implemented in SurfaceTopography

[sannant]

I think the reason it is not implemented is that you need the size of the global array.

Options:

• specify grid size at initialization of Reduction and store it as property (it can then be used similarly as numpy)
• call pnp.sum() / np.prod(nb_grid_pts)

Of course another option is to compute the total nb of entries on the fly in pnp.mean (i.e. making two reductions)

i.e. pnp.mean(array) = pnp.sum(array) / pnp.sum(len(array))