Notes on creating a new survreg.

 The fitting logic has been updated
	1. Estimate the mean model first.  This gives a more solid estimate
of sigma.  
	2. From this point, use the glim trick to go to the full model.
	3. Levenberg-Marquardt has been replaced with a Fisher scoring
step.  The step isn't always as good, but it is always SPD.  With better
starting estimates we rarely use the Fisher anyway.


There are 3 data sets here for which the survival4 version of survreg fails.
   Capacitor: There was an error in my initial values
   Peterson: Very hard -- a factor variable for which some groups have n=2.
   Donnell: The likelihood surface looks like a kidney, so even close to the
solution a slight misstep lands in a region for which the information
matrix is not positive definite.

  The capacitor data now works like a charm.

  The Peterson data takes 8 iterations, but makes it.  There is some
step halving in the sequence because of near singularity.

  The Donnell data fits in 9 iterations.  A good initial estimate along
with a decent first step (scoring based) are the keys.

  In all of these, censorReg gets to the solution quicker, e.g. 5 iterations
for the Donnell data.  I expect that there are ill-conditioned data sets
for which survreg won't converge by censorReg will.  (If you find one,
send it to me however.  I'd like to add it to the challenge list).  Survreg
has more flexibility, though, particularly with respect to penalized
models.