Overflow in scheduler_reweight

The current version of master overflows in scheduler_reweight() and leads to a crash.

This follows changes I made when changing the gravity calculation. I made some of the variables floats instead of integers. We now crash (with -e) when converting these floats to integers. That happens with MPI switched on as the communication costs of these is set to 1e9. The total cost then goes beyond the range of int and we crash with and FPE when converting back to int.

So, did we expect this behaviour when using only int? Do we currently rely on integer over-flow to set the task cost and hence do domain decomposition?

Should this be changed? Or should I roll back my changes (set everything back to int)?

@pdraper do you have any opinions on this?