perf(MPIArray): Write redistributed chunks directly into target array

[ljgray]

Reduce memory usage and remove an extra step from redistribute. This offers some performance improvements when the number of ranks is small but MPI overhead dominates at larger scales.

[jrs65]

Great!

There's another memory optimisation I've been thinking about for a while which would get rid of the full third copy (contained within buffers). We could do the message passing in stages, i.e. round 1 each rank passes to its (cyclic) neighbour, waits on the results and moves into the final location, round two it passes to its next-but-one neighbour, etc. Then you don't need the full array buffer. It may be a bit slower in terms of the network transfer, but these redistributions are one of the biggest unpredictable eaters of memory, and if we can get them down we can probably run on fewer nodes at once.