Sensitivity of LM minimiser to missing data

[Ian Culverwell]

Some questions about the correct treatment of missing data in the LM solver arose during the development of the differenced bangle solver in #700. In particular:

1. Is masking the data via

   delta_y = (y%bangle - obs%bangle) * y%weights
   

valid if the R matrix is non-diagonal, e.g. for refractivities;

2. Why don't we need to mask the K matrix fore refrac and bangle, so that parts of it relating to missing obs make no contribution to the error matrix A, as we needed to do for differenced bending angles in order to get a match to the original source code, thus:

     DO i = 1, nstate
       K(:, i) = K(:, i) * y%weights(:)
     ENDDO
   

This ticket is the place to record the results of investigations.