Better handling of non-full-rank model specifications with predict.BTm/BTabilities

[hturner]

Relates to https://bugs.r-project.org/bugzilla3/show_bug.cgi?id=16158

The least that should happen is for a warning to be issued, as predict.lm currently does, when "newdata" is used on a model with non-full-rank specification.

We could perhaps go still further, as Russ Lenth suggests in the above bug report, ie check the estimability of each parameter combination that's being asked for in "predict".

The package written by Russ Lenth may be an alternative to our checkEstimability function.

A simple example is this: standard Bradley-Terry model with 4 players, and match results such as
A beats B 10 times out of 40
C beats D 30 times out of 40
Here there's no information to allow prediction of, for example A vs C. But predict() gives an estimate and st error for such, and since it's nonsense there ought to be a warning accompanying them. (or ideally the result should be NA).