Call bulk_flux only in the first run of the nonlinear model

[arango]

The data assimilation algorithms were updated to call only the routine bulk_flux in the first call to the nonlinear model when both BULK_FLUXES and NL_BULK_FLUXES are activated. The surface fluxes computed from bulk_flux are stored in the basic state NetCDF file. These are processed in the outer and inner loops and adjusted with the minimization increments.

Many thank to Andy more with his help testing this option.

I did some cleaning:

• Renamed is4dvar_lanczos_ocean.h to is4dvar_ocean.h
• Inserted cgradient_lanczos.h into cgradient.F
• Removed obsolete cgradient_lanczos.h
• Removed obsolete driver grad_ocean.h and its associated C-preprocessing option GRADIENT_CHECK