Corrected a couple of bugs in the W4DPSAS driver:

- Remove the writing of the nonlinear model initial conditions at the end of the inner loop in w4dpsas_ocean.h. The correction due to data assimilation is added in main3d.F to the first guess state. The nonlinear model is always started from the first guess. Otherwise, the solution dynamical range with accumulates in each outer loop. The final (optimal) nonlinear initial conditions are written after done with data assimilation.
- Modified def_rst.F, wrt_rst.F, tl_def_his.F, and tl_wrt_his.F to write the basic state vertical mixing coefficients when any of the vertical parameterization are activated. This will allow to control instabilities due to data assimilation. Recall that the finite amplitude tangent linear model (RPM) is used as the basic state for the next outer loop iteration in the W4DVAR algorithm. Since we are not using the tangent linear of these vertical parameterizations, the nonlinear model values are a good approximation in the Picard iterations.