[Developers] incomplete beta function is broken

[dave fournier]

The problem that I am working on is dealing with observations at or very 
near 0 or 1
in a beta regression.  One solution is to change the observations to 
something
very near the boundary -- but how close? what seems to be more elegant is to
change the model to group the observations near the boundary and use
the cumulative distribution for the beta. that involves
stuff like

         1.0-betai(1-e)

for some small value of e like .01 or .001 etc.

When I built the model it did not run very well. In the inner optimization
the newton-rapshon steps did not quite reduce the gradient enough.

I have now recompiled the code using the old version of betai
and it runs much better. I think that the new betai is not quite smooth
enough for some reason.     To see what the problem might be
consider a function that looks like

/\/\/\/\/\/\/\/\/\/\/\

Now if the little saw tooths are very small the function is very close to
a constant, but the derivatives are not close to 0. It is possible that the
new betai has problems like that which are so small they are difficult 
to see
in isolation, but affect the performance of the ADMB code.