Corrected model/observations statistics

[arango]

We found a logic problem in stats_modobs.F when computing the statistics between model and observations. The logic for obs_scale was wrong when there are more survey data that needed in analysis time-window. I added a _FillValue of zero to this NetCDF variable. Also, the _FillValue for the model variables was also changed from 1.0E-6 to 1.0E+35. I modified this routine to use the new generic NetCDF interface. Many thanks to Gordon Zhang for bringing this to my attention.

I also corrected the outer and inner loop variables in the 4DVAR drivers to use local variables (my_outer and my_inner) in the actual loops. This is because Fortran adds 1 to the couter to get out of the loops. We need the global variables to have actual values.