[phase closure] sum closure memory opt [ci skip]

[basaks]

memory optimisations during sum closure

[richardt94]

I pushed a commit which changes this to a capital R reduce. This fixes further issues that lower-case reduce was having with creating a 13 GB pickle object to communicate its part of the ifgs_breach_count_arr back to the root node. Using Reduce is safe because while mpiops.array_split may result in different numbers of loops being sent to each process, the accumulated breach count arrays are all numpy arrays of the same size (height of ifg, width of ifg, # of ifgs) so we can use preallocated buffers for our reduction operation (and the sum over axis 0 is default behaviour of MPI.SUM so we don't need to mess with a custom sum0_op). Still a couple of things to change before we can merge this:

1. I'm not sure if it's safe to use the same buffer for sending and receiving for the reduction - if not we'll need to create an empty array of zeros
2. This commit will break the non-MPI option, because I had to directly import mpi4py in sum_closure.py to get MPI.SUM. Should be easy to fix this by exporting MPI.SUM out of mpiops.

[basaks]

good pick up on the Reduce use here Richard. Thank you.

[richardt94]

Thanks Sudipta for your work debugging this PR and fixing up the buffer-based Reduce method. As far as I can tell it is now fully functional, so I'm happy to merge it now - might be worth getting @mcgarth 's opinion first