clearify semantics of preconditioner pre/apply/post

I always assumed that for every Preconditioner object, it is necessary to call pre before apply. Then it is safe to call apply as often as necessary to converge and the post ensures cleanup of temporary data.

The pthread tests in dune/istl/paamg/test now also test variants, where the pre is called before cloning the AMG and the running it in several threads. I'm surprised by this behaviour, and it tests a stronger requirement than I always assumed.

What is the proper behaviour here?