Corrected several bug in the 4DVar algorithms

[arango]

Several bugs were corrected in the 4DVar algorithms when running with multiple outer loops. This included the computation of the cost function. We also corrected the MPI exchanges in the adjoint convolution (ad_convolution.F and ad_variability.F) to follow adjoint rules. This was not really a bug now because the values in the ghost point were zero any way. However, this may not the case in the future.

Many thanks to Andy Moore for reporting and correcting these problems.