[bugfix] Always broadcast NNCs from input.

[blattms]

This was only done if there were NNCs from aquifers, but it should always be done whenever there are NNCs.

Those NNCs also hold information of (additional) transmissibilities for a grid intersection. While the intersection is already represented by the grid, the transmissibility is not. This PR fixes this and makes the NNcs available on all processes.

If I missed where we broadcast NNCs, then please tell me.

[blattms]

jenkins build this please

[blattms]

We must be doing this somewhere else already. Just need to find it.

[akva2]

These are already broadcast during the EclipseState broadcast, ie, https://github.com/OPM/opm-common/blob/master/opm/input/eclipse/EclipseState/EclipseState.hpp#L151

[blattms]

And that seems to happen in readDeck.cpp, which is really early. Why might just want to move that to a later stage. It seem that we don't need it before we have a distributed grid, anyway. At that stage we might also be able to do much smarter things and save bandwidth and memory ...

[atgeirr]

A somewhat common way to do arbitrary non-cornerpoint grids in OPM Flow is to make a disjointed grid with all cells in a row, then adding every connection as an explicit NNC. Will this work make that scale worse in parallel than before?