[Rd] drop.unused.levels

[Terry Therneau]

 A user sent me a query on survreg, which when boiled down is a request
to drop unused levels when setting up the X matrix.
 I don't have a strong opinion one way or the other, but am loath to
make the change: I expect that code somewhere will break, perhaps a lot,
when the length of the coefficients changes.

 On the other hand, at some point this change was made to lm().  Does
the R core have an opinion on which is better, or a pointer to
discussion that underlied the change to lm()?  It could be enough to
push me over the edge.  It would affect survreg, coxph, survfit,
survdiff, coxme, ...

Terry Therneau