Make Norm communication asynchronous

[albestro]

Norm is one of the few points left with synchronous communication. In particular, it uses an MPI_Reduce of a single value, which is a quite special use-case since all our helpers/wrappers/functions deal with tiles instead of simple values.

Instead of making the simple value appear as a Tile, which IMHO sounds a bit cumbersome, I opted for the opposite strategy which sounds more straightforward: implement an asynchronous reduction for the single value use-case.

Main changes:

• now the algorithm is asynchronous and it returns a sender; value is valid just on the root rank as before.
• local reduction does not need to be synchronised anymore (now it is passed asynchronously to the distributed reduction final step)

It might be argued that:

• we might want to make the reduction for simple value available for more algorithms: I would leave it for later in case there will be need of that;
• we might want to keep the norm algorithm returning a value instead of a sender: as first choice I opted for changing it, but open to comments.

[msimberg]

Thanks @albestro!

ETI question still open, but not a blocker for this PR (if ever).

[albestro]

cscs-ci run