Issue when calculating nll for compositional data

[GiancarloMCorrea]

We have discussed this over email, but I wanted to make it formal.

There is an issue calculating the nll for compositional data when the predicted paa (or pal, or caal in growth branch) for a given year and fleet has a small number in it (specifically, the issue arises when a number < 1e-16 is present).

For example, for the ddirichlet function, the issue happens here, calculating log(sx) when sx = 0 so generating NaN.

Likewise, I suspect that something similar happens here, calculating p/(1-pUsed) when pUsed = 1 so generating NaN.

Not sure how to fix this; the squeeze function might help.

[timjmiller]

This issue has been resolved on devel branch.