perf: Start using cells for bonding calculation

[rprospero]

This uses cells to split up the bonding calculation. Instead of parallelising across atoms, we run in parallel across neighbouring cell pairs, which isn't quite as efficient, but hopefully the cell splitting makes up for the loss. One place to pay extra special attention is in line 58 of calculateBondingNode.cpp. I have been doing a dynamic calculation of the size of the box based on the inscribed radius, but this was somehow creating 1000 cells on linux and 2³² cells on Mac. I've now hard coded the cell size, but I would appreciate suggestions on a better way of handling this. At the time I'm writing this, the Mac build is failing due to a server failure for one of our dependencies, so I cannot check whether or not the hard coded size fixed the issue.

To do this, I've added an extra method to cell array that gives all the pairs along with all of the self pairings, which was behind the bug I was dealing with in the standup this morning.

There's also a change switching cells to just store an AtomBase pointer and then adjusting the existing call sites for CellArray.

As future work, we could write a cell atom-pair iterator that would allow us to iterate through the atom pairs in same or neighbouring cells. This would both simplify the code (likely in multiple places) and prevent waiting on the thread with the most populated cell to finish.