VERY IMPORTANT: Updated Saddle Point 4D-Var

[arango]

Continue improving the saddle-point 4D-Var to solve latency issues when writing and reading NetCDF data.

• Introduced SKI_REORTHO to skip the re-orthonormalization of the Arnoldi vectors in gmres.F. In some applications, it does not affect convergence or results, but it accelerates the computations. It can be done because GMRES minimization is a very robust method.

• Added nf90_share to netcdf_create and netcdf_open to facilitate access to a NetCDF written and read by different processes. It is only avaluable for the NetCDF-3 library.

• A new logical argument, Lclose, is added to def_state.F to close the file after being created to alleviate access latency between split sub-groups of PETs.

• A call to netcdf_sync is added to the beginning state_read.F. It Synchronizes NetCDF data to disk. It allows processes open for reading data to update the number of records written by other processes. The writer flushes buffers to disk, and the reader makes sure that it is reading from disk rather than from previously cached buffers.

• Reworked action=3, in atimesv.F to avoid reading TLmodVal from disk. A collected MPI communication is done in saddle_point.h instead.