Skip nonlinear term communication for forbidden regions

[David Dickinson]

Currently we calculate the nonlinear term for every theta, species, energy and pitch angle point. This includes all the forbidden regions. At high theta resolution the forbidden region accounts for roughly 50% of all data. If we excluded this from the nonlinear term calculation then we could reduce the amount of data communicated by up to a factor of two and can reduce the number of Fourier transforms by the same factor. This could therefore save up to a factor two from the cost of calculating the nonlinear term.

We may be able to save a factor two in the communication from g_lo to the collisions layout by again not communicating the forbidden region.

See PR #549 and PR #547 for related changes.

[David Dickinson]

We could seek to eliminate the forbidden region from the data stored by the relevant layouts, although this may prove complicated. A simpler approach may be to simply identify forbidden points and exclude them from the redistribute and possibly the work involved with these points. The latter option is simpler but will result in load imbalance as we would be allocating some processors less work than others (as some will get more forbidden regions than others).