psMinimizeLMM.c handling of fitting failures

[eugene]

the code using the LUD inversion was resulting in NaN parameters, indicating
failed solutions. this is not a very good solution to the fitting. It is
better to detect the singular matrix and return an error. I have reverted to
the GaussJordan matrix inversion which returns a status condition if the matrix
in not invertable. This is now caught (in SetABP) and is used by psMinimizeLMM
to modify the guess (increase lambda) or to eventually result in a
non-convergence. Since this is a valid operating condition (just a pooly
constrained set of data), I have changed the psError messages which were related
to these states to psTrace messages.

changes in CVS HEAD

[Paul Price]

Fixed by EAM.