Use request-based RMA

[matt-frey]

Instead of MPI_Get/MPI_Put we use MPI_RGet/MPI_RPut. A MPI_R*` routine allocates a communication request object that can be use to wait for completion. See for example https://www.open-mpi.org/doc/v4.1/man3/MPI_Rget.3.php

[matt-frey]

We have now compared almost 23.5 billion merges. The parallel version returned the same results as the serial version.

Test 1

Total number of samples: 12000
Seeds: [45, 89, 351, 874]
MPI ranks: [16, 32, 64]
Number of parcels per sample: 1310720
Number of parcels per cell: 40
Number of fails: 0
Number of merges: 5802331176

Test 2 with 16, 32 and 64 MPI ranks; seed 875

#samples, #fails, #merges: 13200 0 6382621749

Test 3 with with 16, 32 and 64 MPI ranks; seed 43

#samples, #fails, #merges: 11800 0 5705727073

Test 4 with 72, 96 and 120 MPI ranks; seeds 45 89 351 874

#samples, #fails, #merges: 11500 0 5560559699

[sjboeing]

Just had a quick look. Happy for you to merge this once you have run a test case (like RT or moist bubble).