Process indices in ComputeIndexOwner by intervals

[kronbichler]

This PR builds on top of #8772, and only the last commit actually implements something new. The new functionality is:

1. We process the owner of the dictionary section in the compute index owner code interval by interval, not dof by dof
2. A minimal grain size has been introduced.

Fixes #8785.
Fixes #8293.

[kronbichler]

Just as a remark, the contribution with the grain size is quite important on many ranks. Taking the example listed in #8293 (see there for the labels) with 49k MPI ranks, here is the difference between this patch applied and not.

Before:

With this patch:

Compare e.g. the mg transfer data point where we call Utilities::MPI::compute_index_owner() as well as set up two Utilities::MPI::Partitioner objects for each level. What happens: On level 2 we have 256 cells and, due to FE_Q<3>(5), 35301 DoFs. In the old setting, we would spread these indices onto a dictionary with 1 DoF each (and some MPI ranks do not get any index). As only 256 out of these 35k ranks actually own cells on that level, this is essentially a one-to-all communication that hurts on that scale. With the grain size, this problem disappears.