reduce number of prepare_coarsening_and_refinement() calls

reduce number of prepare_coarsening_and_refinement() calls

I noticed that we call prepare_coarsening_and_refinement() many times
when using a p::d::Triangulation. Many of these are certainly
unnecessary:

1. distributed tria for mesh_reconstruction_after_repartitioning. This
is used to remove all refinement and smoothing will be called inside the
execute...() call below. No need to call our more expensive version.

2. serial execute_coarsening_and_refinement() should not call the
derived version, which in turn calls the serial version.

For the test tests/distributed_grids/refine_periodic_01:
Before: 154 sequential, 154 parallel
After: 126 sequential, 53 parallel

fixup