Consensus algorithm: answer needs to be of known size

[bangerth]

In the ConsensusAlgorithms interface, the prepare_buffer_for_answer() function exists because we want to post an MPI_Irecv operation right after sending the request to another processor. A downside of this interface is that we need to know the size of the answer message, unlike for the request which can be of variable length. This is a bit awkward because it doesn't allow sending variably sized messages: the receiver of the answer needs to already know the length of the answer.

A possibly better interface would do away with the prepare_buffer_for_answer() function by not posting the MPI_Irecv right away and checking its completion later on (in clean_up_and_end_communication()), but to work through a loop of regular MPI_Probe/MPI_Recv functions in clean_up_and_end_communication() for all incoming answers. We can do this because we know how many answers we expect, so we can just loop as many times as we expect answers; the Probe then gives us the message length, we can resize the answer buffer, and then put the answer into it in the MPI_Recv.

[bangerth]

Let me reopen this to remind myself that I need to document that the callback is no longer used.

[bangerth]

Fixed now with #13267.